EP4292101A1 - Individualized medical intervention planning - Google Patents

Abstract

The invention is directed at a method of operating a decision support system for evaluating a medical intervention plan for the medical intervention of a disorder in a human or animal body, the method comprising the steps of: obtaining, by the decision support system, patient data indicative of one or more patient characteristics for an individual patient; obtaining, by the decision support system, medical intervention data indicative of the medical intervention plan for the medical intervention; identifying, based on the medical intervention data, at least one prediction model associated with the medical intervention plan; applying, based on the patient data, at least one modifier model for modifying the at least one prediction model based on the patient characteristics of the individual patient, such as to obtain a modified prediction model; and determining, based on the patient data and using the modified prediction model, a medical intervention outcome in terms of one or more probability values associated with one or more health status events indicative of medical intervention result probabilities. This document further describes a decision support system applying the method.

Description

Title: Individualized medical intervention planning

Field of the invention

The present invention is directed at a method of operating a decision support system for evaluating a medical intervention plan for the medical intervention associated with a disorder in a human or animal body, as well as to a decision support system therefor.

Background

Precision medicine is the future of health care. As a technology-intensive and dependent discipline, medicine will be at the vanguard of this impending change. It will impact medical intervention of individual patients, the way in which patients decide about whether and which medical intervention is desired, but also the way clinical trials are done.

However, to bring about precision medicine, a fundamental conundrum must be solved: human cognitive capacity, typically constrained to five variables for decision making in the context of the increasing number of available biomarkers and therapeutic options, is a limiting factor to the realization of precision medicine. Given this level of complexity and the restriction of human decision making, current methods are untenable.

Presently, questions related to screening, vaccination, diagnosis, treatment, recruitment for clinical trials, or follow-up treatments, are typically answered based on generic guidelines, but are not individualized. Some patient decision aids are available, but these are typically “hard-coded recommendations rules”, based on such guidelines. Some decision support systems typically compare variations of a same type of treatment (e.g. radiotherapy with or without spacer, or proton versus photon radiotherapy) while most often there are more options. Presently, no decision support systems are known that enable to compare completely different types of medical intervention with each other, tailored to the characteristics of an individual patient or a group of similar patients.

Furthermore, in present days clinical trials, related to diagnostic- therapeutic or follow-up question, may be performed by identifying a cohort of patients, with suitable characteristics and diagnoses (typically inclusion and exclusion criteria), who are willing to participate. However, with precision medicine this is sub- optimal because the benefits or disadvantages for given patient characteristics are not specifically defined. Thus for certain medical interventions, this may render the group of potential patients too small and the patients themselves too difficult to identify.

Summary of the invention

In this document, the term ‘medical intervention’ is to be interpreted as including any kind of medical intervention associated with any kind of disorder, regardless of whether the disorder or disease is diagnosed with a patient or whether it relates to a potential future disorder that may be developed (such as in respect of vaccination for a potential disease, or screening of patients (regardless of, or initiated due to their present state of health) to identify potential health risks or potentially undiscovered disorders). The term ‘medical interventions’ may thus for example refer to: screening of disease in asymptomatic patients, vaccination against pathogens or chronic disease, diagnosis of diseases in symptomatic patients, treatment of diseases, recruitment in clinical trials (for example trials approved by an institutional review board (IBB)), or a follow-up after treatment or diagnosis.

It is an object of the present invention to overcome these disadvantages and to provide a method and a decision support systems (DSS) that enables to evaluate a medical intervention plan for the medical intervention associated with a disorder in a human or animal body, in particular suitable for individualized medical intervention.

To this end, there is provided herewith a method of operating a decision support system for evaluating a medical intervention plan for the medical intervention associated with a disorder in a human or animal body, the method comprising the steps of: obtaining, by the decision support system, patient data indicative of one or more patient characteristics for an individual patient; obtaining, by the decision support system, medical intervention data indicative of the medical intervention plan for the medical intervention; identifying, based on the medical intervention data, at least one prediction model associated with the medical intervention plan; applying, based on the patient data, at least one modifier model for modifying the at least one prediction model based on the patient characteristics of the individual patient, such as to obtain a modified prediction model; and determining, based on the patient data and using the modified prediction model, a medical intervention outcome in terms of one or more probability values associated with one or more health status events indicative of medical intervention result probabilities. The present invention enables to perform an early evaluation of a projected or potential medical intervention plan, prior to the medical intervention plan being carried out, for individual patients. The method enables to provide a tailored risk assessment of a certain medical intervention plan, by enabling to modify a prediction model on the basis of individual patient characteristics.

For example, suppose a first study provides a risk assessment for a screening with total body MRI in asymptomatic patients. This examination could detect an asymptomatic cancer treated with side effects but could also be costly, generate false positive finding that could generate unnecessary medical examinations, anxiety and decrease of quality of life. In accordance with the present invention, the method enables the decision support system to apply the risk assessment of the first study to an individual patient, while taking into account specific patient characteristics of that patient.

In another example, suppose a first study provides a risk assessment for a treatment of a certain disorder using a specific treatment type, let’s say the palliative treatment of advanced non-squamous cell lung cancer (NSCLC) in general by means of chemotherapy using a certain drug. In accordance with the present invention, the method enables the decision support system to apply the risk assessment of the first study to an individual patient, while taking into account specific patient characteristics of that patient. For example, if it is quantified by a second study that the specific type of tumor with which the individual patient is diagnosed will respond poorly to any type of chemotherapy and that there is a high risk of toxicity and decrease of quality of life, then the risk assessment of the first study can be modified by the decision support system based on the quantification by the second study. The decision support system will then provide an individualized risk assessment for this specific case, enabling the patient to conclude that the side effects of the medical intervention will be more significant in view of the expected diminished benefits.

The above is merely an example, and the invention is not limited to the evaluation of medical intervention plans for certain types of cancer. In principle, any disorder may be evaluated in this way, even regardless of whether an individual patient really suffers from the disorder. Also in principle, any medical intervention may be evaluated in this way, such as screening of diseases in asymptomatic patients, non-compulsory vaccination, diagnosis in symptomatic patients, treatment, recruitment in a clinical trial and follow-up of a disease. Indeed screening strategy and diagnostic plan (e.g. do we need to do a screening with MR for breast in young patients or of pancreas cancer? Do we do an invasive biopsy an expensive imaging or a genetic test that could have side effects (e.g. bleeding during the biopsy) or delay the start of the treatment?) may also be evaluated by this approach. The same is true for follow-up plan after a diagnosis and/or a treatment (how often must the patient should be seen? Should we do regular imaging? Should we advise a digital tool with questionnaire and or sensors? If yes what are the results that should trigger a visit?).

For example, the decision support system enables to evaluate hypothetical cases as well. Suppose, for example, an individual user wants to assess the risk of side effects of vaccination for COVID 19, in that case the user may provide biometric, genetic data, HLA subtype and other relevant information as patient data to the decision support system, and select the application of a particular COVID 19 vaccine as medical intervention plan. The decision support system may access a database in order to obtain a prediction model for the respective vaccine. The prediction model, in accordance with the invention, can be modified based on specific patient characteristics of the user. For example, suppose the user in question is suffering from an auto immune illness or receives a specific medical intervention for a different disorder which impacts his/her immune system or has a HLA type that would decrease the effect of the vaccine, then the impact of this specific patient characteristic state may be quantified to provide a modified prediction model, e.g. on the basis of other studies or an estimation. The decision support system will determine, based on the patient data provided by the user and using the modified prediction model, a medical intervention outcome in terms of probability values associated with one or more health status events indicative of medical intervention result probabilities. In this case, the decision support system may show that the impact of the user’s auto immune illness on the side effects of vaccination for COVID19 will be limited.

For the sake of clarity the patient data can be clinical and/or demographic (e.g. age, body mass index general status, number of steps done per day), biological (e.g. blood tissue biomarkers), genetic (germ cells somatic mutations or polymorphism, HLA haplotype, somatic mutations of cancer cells), (quantitative) imaging data or data based on sensors (e.g. number of step, use of the phone, speed of speech, type of speech, glycemia, blood biomarkers) or digital questionnaire.

In some embodiments, the invention enables to compare two or more different medical intervention plans with each other. The term medical intervention plan is to be interpreted broadly, such as to include for example a ‘do nothing’ strategy. In the above example, this will for example enable the user to evaluate the effects of vaccination for C0VID19 in comparison with doing nothing and attracting an infection with the coronavirus SARS-COV-2, in order to support a decision to take the vaccine. In other words, the method of operating a decision support system in accordance with the present invention enables take away any doubt prior to the medical intervention, in this case vaccination, the advantageous of which, for the individual patient and in the given example to the society, are well understood.

In some embodiments, the method further comprises a step of obtaining preference data indicative of a health status preference or risk profile preference of the patient related to the medical intervention associated with the disorder; and comparing the preference data with the one or more probability values for determining an agreement level therebetween. In this case, the user may evaluate the medical intervention in relation to certain desired personal health state criteria. For example, suppose a certain individual patient is single, has small children and want to decrease the risk of a life threatening disease, some screening examinations could be indicated. As another example, suppose a certain individual patient wants to be able to rely on certain body functions after medical intervention, e.g. in order to be able to continue practicing his/her profession, the maintenance of that body function may be provided as a health criterion, i.e. as preference data. For example, a photographer suffering from a tumor (e.g. an Optic Nerve Sheath meningioma) behind one of his eyes, may prefer to receive a medical intervention which provides — in addition to curing the disorder - a good probability of maintaining sight in both eyes. A piano player suffering from a repetitive strain injury in one of his/her arms will prefer a medical intervention that allows him/her to continue playing. An anxious patient living remotely want to minimize the follow-up visits at the hospital. The decision support system in accordance with these embodiments may take such preferences into account and specifically allows to evaluate the medical intervention in terms of these preferences, e.g. to maximize the probability that the outcome will match the preferences. Providing preferences is not essential and may be dispensed with for other embodiments. However, in those cases where these preferences are to be taken along and may not be available from the patient, they may for example be obtained by anamnesis, captured from family members or predicted based on previously trained models or database of similar patients having given in the past their preference (“imitation learning”). The “static” preferences before the medical interventions can be taken into account but also the dynamic preferences varying with time. The preferences have an impact of the weight of certain models in the prediction of the benefits of the medical intervention but can also play a role in the chosen models (e.g. if there are three toxicity models related to a medical intervention but only one toxicity “x” is relevant for the patient, only the prediction model of toxicity “x” will be taken into account in the baseline DSS).

In some embodiments, the prediction model provides statistics data, wherein the statistics data enable to determine, based on at least one of the patient characteristics, at least one of the probability values associated with at least one of the health status events. The prediction models may be based on any study performed to evaluate a medical intervention or the impact of certain health states. Many of such studies quantify the results based on regression models over the cases concerned during the study. These regression models are typically be provided by nomograms. A nomogram, also called a nomograph, alignment chart, or abaque, is a graphical calculating tool: a two-dimensional diagram designed to allow the approximate graphical computation of a mathematical function. It typically consists of a set of n scales, one for each variable in an equation. Knowing the values of n-1 variables, the value of the unknown variable can be found, or by fixing the values of some variables, the relationship between the unfixed ones can be studied. Therefore, in some of these embodiments, the at least one prediction model is a regression model providing a predictable dependency between the at least one of the patient characteristics and the at least one of the probability values, wherein the method comprises a step of: receiving, by the decision support system, a nomogram wherein the nomogram quantitatively describes the dependency between the at least one of the patient characteristics and the at least one of the probability values; and analyzing the nomogram such as to determine, from the nomogram, coefficients of the dependency. Any suitable kind of analysis may be applied, for example, the nomogram may be analyzed by generating non-real patient characteristics and performing a curve fit over the predicted probability values. This will yield the coefficients which define the prediction model, which can then be stored as prediction model data for use by the decision support system.

These models alternatively or additionally also be provided by digital interface and be fed with input data manually and/or automatically from the personal patient database, phone, electronic health record, GPS tracking device or public database of weather, pollution, allergens, presence of air particles, risk of infection or a terrorist attack (this list is not exhaustive). This allow to continuously and automatically update the models with the new continuously changing variables.

In some embodiments, the modifier model includes modifier data, the modifier data being indicative of an impact of a particular patient characteristic state on the at least one of the probability values associated with the at least one of the health status events. For example, the modifier data may quantify - as described earlier - the impact of the occurrence of a certain biomarker for that patient, or the impact of a specific habit (e.g. smoking). In some embodiments, the modifier data includes a hazard ratio associated with the particular patient characteristic state, such as a hazard ratio associated with a biomarker (of any type, e.g. biological, genetic, imaging, volatile organic compounds, etc.). A hazard ratio provides the ratio of the hazard rates corresponding to the conditions described by two levels of an explanatory variable. For example, if a population in which the biomarker occurs suffers from a certain side effect at twice the rate per unit time as a control population wherein the biomarker does not occur, then the hazard ratio would be 2, indicating higher hazard of the occurrence of the side effect from the medical intervention for patients with that biomarker. These hazard ratios can be used as modifier data to modify the prediction model.

In some embodiments, the modifier model includes one or more further prediction models different from the at least one prediction model associated with the at least one medical intervention plan, the further prediction models enabling to modify the at least one prediction model such as by averaging the probability values or correcting for occurrence of a particular patient characteristic state. As referred to above, prediction models evaluating a same medical intervention may be combined in this manner to provide a modified prediction model to increase the accuracy or allow to correct for specific cases occurring in one of the cohorts concerned in each study. This can be done by averaging the results or adding the results and performing a new regression over the combined groups. There are several ways in which the two prediction models may be combined in order to provide a modified prediction model in accordance with these embodiments.

In some embodiments, the patient characteristics include one or more elements of a group comprising: demographic variables, such as gender, age; one or more body parameters, such as body size, weight, body mass index, general status or fat percentage; genetic characteristics, such as gene polymorphism, human leukocyte antigen, haplotype or somatic or germ-cell mutations; co-morbidities; one or more body- or physiological characteristics, such as hair color, skin type, speed of speech, type of speech, number of steps done per day, blood type; biological age; imaging data; data acquired from sensors (e.g. number of step, use of the phone, ); user input; medical status data, such as medications taken, illness progression status, allergic response data, presence of a biomarker, illness data (comorbidity); medical history data, such as medication history, earlier received medical interventions; behavioral data, such as use of social media and phone, practiced sports, physical exercise frequency, smoking habits; location data (automated through link with a geolocalisation/GPS tracking device or not); climate or weather data, such as temperature, humidity, precipitation, pollution, allergens, air particles; environmental or social data, such as risk of infection, occurrence of an epidemy or pandemic, risk of terrorist attacks.

Furthermore, in some embodiments, the medical intervention data include one or more elements of a group comprising: indication of a medical intervention type, such as medical intervention by surgery, interventional radiology (e.g. radiofrequency ablation, microwave ablation,), drug therapy (such as chemotherapy, immunotherapy, targeted agents, hormonotherapy), irradiation therapy or radiotherapy, tumour- treating fields ((TTFields), also known as alternating electric field therapy), movement therapy; medical intervention specifics, such as medicine data of a particular type of medicine to be applied, a data relating to particular type of surgery to be performed; data relating to enrichment of the microbiome, data relating to treatment with a genetically modified organism (virus or bacteria), data relating to treatment with stem cells, data relating to geroprotectors. A main advantage of the present invention is that it enables to perform an individualized evaluation of a wide range of different modalities of medical intervention, and is not restricted to different versions of a same type of treatment of medical intervention.

In some, other or further embodiments, the health status events include one or more elements of a group comprising: cured; recurrence of the disorder after a predetermined time; decreased or accelerated progression of the disorder; occurrence of post-medical intervention adverse effects, such as a medical intervention related disorder or complication, medical intervention related damage to the body, loss of body function; quality of life, depression, happiness, death.

In some embodiments, the step of determining a medical intervention outcome determining a health state score based on the probability values associated with the one or more health status events. The health state score may for example be a number between 0 and 10 - wherein 0 relates to death and 10 relates to completely cured with no side effects. Alternatively, though, the decision support system may also provide a risk assessment in terms of probability values, for example: 36% probability of suffering from side effect A, 23% probability of suffering from side effect B, 83% probability of curing the disorder.

In some preferred embodiments, the method comprises comparing a first medical intervention with a second medical intervention for the medical intervention of the disorder in the human or animal body, wherein the step of obtaining the medical intervention data comprises: obtaining first medical intervention data indicative of the medical intervention plan for the first medical intervention, and obtaining second medical intervention data indicative of the medical intervention plan for the second medical intervention; wherein the step of identifying the at least one prediction model comprises: identifying at least one first prediction model for the first medical intervention and identifying at least one second prediction model for the second medical intervention; wherein the step of applying the at least one modifier model is performed for at least one of the first or second prediction model; and the step of determining the medical intervention outcome comprises determining a first medical intervention outcome based on the first and the second prediction model, as modified by the at least one modifier model, such as to yield a first and a second medical intervention outcome; wherein the step of comparing the first medical intervention with the second medical intervention is performed based on the first and a second medical intervention outcome. These embodiments have briefly be described above in relation to vaccination for COVID 19, but may of course be applied for any two or more medical interventions in relation to any disorder. For example, this enables to evaluate and effectively select a preferred diagnostic path, e.g heavier with MRI of the brain, for diagnosing the extension of certain form of lung cancer more likely to develop brain metastasis, or compare the results of surgery with those of irradiation, chemotherapy or immunotherapy, or compare a standardized follow-up (e.g. visit every 4 months) with an adapted one (e.g. visit every 2 months or use of a digital app to monitor symptoms and quality of life, if visit are likely to create anxiety) or use or regular brain imaging). As a further example, this enables to evaluate and effectively select a preferred medical intervention for treating a certain form of cancer, or compare the results of surgery with those of irradiation, chemotherapy or immunotherapy.

In some embodiments, wherein the method comprises a step of obtaining preference data, the method comprises performing multiple times during a time period the method comprises the steps of: identifying the medical intervention model, applying the modifier model and determining the medical intervention outcome; and at least one of the steps of: obtaining patient data, obtaining preference data or obtaining medical intervention data, for enabling to capture changes in the data over time. In particular, it is advantageous to perform the step of obtaining preference data over the course of time, as these preferences may shift dependent on the patients age or phase of life. For example, for a parent having children, the preferences may become different once these children no longer live with the parent. Also, the preferences of a person exercising a certain profession may change after retirement. A dynamic method for a dynamic decision support system enables to capture such changes and re-evaluate periodically or upon detecting a change.

Certain embodiments of the above apply Reinforced Learning (RL) by using interaction experience with the reality and an sequential evaluative feedback to implement the above in a dynamic setting. For the sake of clarity RL includes model- based (or planning) methods and model-free (or learning) methods. Medical interventions (e.g. screening procedures or treatment regimes) are usually characterized by a prolonged and sequential procedure. Therefore one DSS before the medical intervention is not sufficient what is needed is a “dynamic DSS”. Therefore the prediction of the model continuously improve based on real world events pre- classified as reward (e.g. the quality of life is excellent, the patient has a good urinary function), a positive value function V, and/or penalties (e.g. a side effects, a relapse of the disease) a negative value function V. Those reward and penalties modelling the improvement of the models can be determined by expert knowledge.

In some embodiments, the weight of the penalties and/or reward of the models can be modified by the preference of the patient or, one step further, the relevant penalties (e.g. toxicity “x” not toxicity “y”) and reward (e.g. quality of life over quantity of life) can be chosen by the patients. This principle can be applied in two situation: a) Before the medical intervention: When a databases is available a new individualized RL-based models can be created with specific penalties and/or rewards based on the preference of the patient. B) During the medical intervention: In case of long term medical intervention such as a treatment of a specific patient or group of patients, the models can be adapted based on the updated RL-based models taking into account the recent events and the updated preference. We can speak of a “dynamic, preference-based individualized DSS”. As extreme example, for the purpose of illustration, a patient treated for an operated prostate cancer treatment receives adjuvant hormonotherapy which cause sides effects (fatigue, hot flashes, bone thinning (which can lead to broken bones), loss of muscle). Having some financial responsibilities toward his son completing his academic studies, he decided initially to maximize his probability of cure. Two year later, the side-effects are still present, he has had a hip fracture, his son is financially autonomous and consequently his preferences have changed towards prioritizing his quality of life over quantity of life. The individualized RL-based models propose to stop the adjuvant hormonotherapy: this should lead to an improve Quality of life of 60% and a decrease of biochemical free survival of 7%. For the sake of clarity the approach described in “b” can also be used for a group of patients or within an in silico trial by a legal entity such as a department, a pharmaceutical company, a clinical trial organization, an hospital or a government as a way to optimize the treatment based on preferences-based reward and/or penalties.

In some embodiments, of the method described in accordance with the invention, the step of obtaining, by the decision support system, patient data comprises: generating, by a controller, patient data for a plurality of individual virtual patients, wherein each individual virtual patient includes one or more patient characteristics associated therewith, the plurality of individual virtual patients thereby forming a virtual cohort. This particular class of embodiments allows to perform an in silico clinical trial based on virtually generated patient data or reusing real data of patients treated in the past with validated Al -based models.

Brief description of the drawings

The invention will further be elucidated by description of some specific embodiments thereof, making reference to the attached drawings. The detailed description provides examples of possible implementations of the invention, but is not to be regarded as describing the only embodiments falling under the scope. The scope of the invention is defined in the claims, and the description is to be regarded as illustrative without being restrictive on the invention. In the drawings:

Figure 1 is a schematic illustration of a decision support system in accordance with an embodiment;

Figure 2 is a schematic illustration of a method of operating a decision support system, in accordance with an embodiment;

Figure 3 is a schematic illustration of a method of comparing medical interventions in accordance with an embodiment;

Figure 4 schematically illustrates a challenge encountered with performing a clinical trial amongst a small population of patients;

Figure 5 is a schematic illustration of a method for performing in silico trials using a decision support system, in accordance with an embodiment;

Figure 6 is a schematic illustration of a method of obtaining prediction model data from clinical studies, useable as part of a method in accordance with an embodiment of the present invention

Figure 7 schematically shows a summary of a Markov model for an exemplary case study;

Figure 8 schematically depicts graphs showing distributions of the patient population over different health states over time;

Figure 9 shows a nomogram that may be analyzed in an embodiment of the invention;

Figure 10 shows a fitted graph obtained using a regression model for the nomograph of figure 9.

Detailed description

In this document, the term ‘medical intervention’ is to be interpreted as including any kind of medical intervention associated with any kind of disorder, regardless of whether the disorder or disease is diagnosed with a patient or whether it relates to a potential future disorder that may be developed (such as in respect of vaccination for a potential disease, or screening of patients (regardless of, or initiated due to their present state of health) to identify potential health risks or potentially undiscovered disorders). The term ‘medical interventions’ may thus for example refer to: screening of disease in asymptomatic patients, vaccination against pathogens, diagnosis of diseases in symptomatic patients, treatment of diseases, recruitment in clinical trials (for example trials approved by an institutional review board (IBB)), or a follow-up after treatment.

Without any of the below examples being limitative on the invention, and merely provided to be illustrative, the fields of screening, vaccination, treatment, clinical trial preparation or recruitment, or follow-up treatment could relate for example to the following questions. Screening could for example include the seeking of answers to questions such as: “what is the expected benefit of a control of the glycemia or a screening colonoscopy?” or “what is the probability of a silent cancer?”. Vaccination strategy could for example include the seeking of answers to questions such as: “should a patient receive this vaccine?”, “what will be the level of protection?”, or “what are the side effects and the risk to generate an off-target auto-immune response?”. Diagnosis could for example include the seeking of answers to questions such as: “which examination are preferred when a lung cancer or a diabetes is suspected in a symptomatic patient?”. Treatment could for example include the seeking of answers to questions such as: “should this disease best be treated by medication, surgery or watchful policy?”. Inclusion in a clinical trials could for example include the seeking of answers to questions such as: “what are for this particular patient the benefits and the risks of inclusion in a clinical trial?”. Follow-up treatment of a disease could for example include the seeking of answers to questions such as: “after the treatment of a cancer or a COPD how often should the patent come back and which examination should be done? Should an app for recording metrics by sensors be used, or are questionnaires preferred?’.

Figure 1 schematically illustrates a decision support system 1 in accordance with an embodiment of the present invention. In the system 1, a decision support server 3 including a controller 5 and an internal or external memory 6 forms a central part of the system 1. The decision support server 3 is connectable to a data communication network 8, which may be a wide area network or local area network or else. Through the data communication network 8, the decision support server 3 is communicatively connected to a plurality of data repositories 10. The various databases or data repositories are optional, and in principle may also be available in the internal or external memory 6. Likewise, instead of using the internal or external memory 6 for obtaining or storing data, any of the data used in the method of the present invention may also be obtained from or stored in any of the databases 10, through the data communication network 8. Hereinafter, reference will be made to database 10, but it is to be noted that any number or type of database may be applied either locally or remotely from the decision support server 3.

Some of the data obtained by the decision support server 3 may be provided as input through any type of input means, such as keyboard 7 or directly from diagnostics equipment 9 such as a CT scanner 9. Furthermore, the decision support system 3 may be connected to a hospital management system (not shown) or other system comprising patient health information to retrieve data therefrom. In the illustration of figure 1, in accordance with one embodiment, user data may be obtained from a patient/user via a mobile telephone 15 connectable through a wireless link 13 at a base station 12 to a data communication network 8. For example, the user may provide patient data via a user interface 16 on his mobile phone. The user for example may provide input via an application running on his phone 15. Such an application may enable the user to encrypt his personal data such as to safely provide it to the decision support system 3, and prevent the data from being compromised. Alternatively or additionally, the phone 15 may comprise sensors (e.g. a heartbeat sensor, temperature sensor, or other arbitrary sensor) that enables to capture (live) patient data. Such data captured via sensors may be captured via the application mentioned above or in another manner, and provided to the decision support system 3.

In addition to patient data indicative of patient characteristics such as biometric information, information about the patient such as age genetic profile, HLA haplotype, or preferences, and information about the disorder to be treated (e.g. tumor genetics or imaging data), the decision support system 1 may further require medical intervention data. The medical intervention data defines the specifics of the medical intervention to be evaluated, and for example may identify a type of screening method, diagnostic algorithm, facultative vaccination, medicine and a prescribed dose, manner of admission and frequency, specifics of a clinical trial or a certain follow-up treatment protocol. Alternatively or additionally, the medical intervention type or medical intervention data may be indicative of a screening procedure (for example total body MRI, CT of the pancreas), or a surgical medical intervention to be performed (for example the removal of a part of the pancreas or the removal of a gall bladder). Furthermore the medical intervention data may be indicative of a specific vaccination with a vaccine having a lower risk of off-target autoimmune response, an immunotherapy medical intervention to be performed for a certain type of cancer. In principle any type of medical intervention (e.g. screening, vaccination, diagnosis, trial inclusion, treatment, follow-up) may be indicated by the medical intervention data. Furthermore the invention is not limited to the medical intervention of tumors, but may more generally be applied to medical intervention of many kinds of disorders. For example, the invention may be applied to support the medical intervention of screening of hypertension with a sensor; a case of covid- 19, or a medical intervention of a repetitive strain injury, a trauma, or a bacterial infection; an inclusion in a IRB approved clinical trial with a app-based trial patient decision aids or a follow-up with a digital health app in which a model is used to decide when the patient should be seen by a doctor and at which interval. As may be appreciated, a medical intervention data may therefore also be indicative of, for example, a type of motion treatment to be performed by physiotherapy . The invention is not limited to these types of medical intervention, and any other medical intervention type may be considered.

Furthermore, also the level of detail may vary. The method may be applied to evaluate medical intervention types that are described very detailed, such as screening for breast cancer with Magnetic resonance imaging and genetic test rather than standard mammography such as the admission of a particular drug in a particular admission scheme and dose distribution to cure a certain disease occurring in a particular situation (e.g. as a complication to another disorder). However, the method may also be applied to evaluate medical interventions at a high level, such as comparing the coverage and side effects of two vaccines (based on epitope hotspot, homology of protein vaccine with protein of human proteome, HLA profile of a population); such as comparing radiation treatment versus a surgical intervention for tumors occurring in a certain region of the body, or comparing radiation treatment versus surgical intervention in general.

In addition to the patient data and medical intervention data described above, the method of the present invention further relies on the application of prediction model data which is indicative of prediction models which enable to predict a certain outcome of the medical intervention. For example, a prediction model may be indicative of predicting a probability of survival, a probability of curing the disease, but also a probability of the occurrence of an undesired side effect. Furthermore, the prediction model may be indicative of the chance of diminishing or losing a certain body function, for example paralysis of a limp. It is to be understood that these prediction models are typically based on clinical studies that are performed and published in the field. For example extensive clinical studies are to be performed prior to allowing access and registration of medication for specific diseases. Furthermore, clinical studies are also performed to re-evaluate certain medical interventions already practiced or to compare them with alternative medical interventions. The prediction model data may typically be obtained from database 10 of decision support system 1. Furthermore, the method uses modifier data that enables to modify the prediction model data in order to for example provide a more accurate prediction or to correct for the occurrence or existence of a specific patient characteristic (such as a biomarker, HLA type, proteomic profile and homology with epitopes of a vaccine or a certain patient habit). These modifier data may also be obtained from clinical studies and case studies such as the prediction model data mentioned hereinbefore. The modifier data may also be obtained from a clinical study targeted at studying the specific effect of a certain biomarker or characteristic on the curing of a disease or on the chance of complications. This data may also be obtained from the database 10.

Figure 2 schematically illustrates a method in accordance with the present invention. In step 20, the patient characteristics or patient data is obtained for example from a database 10 or from user input, e.g. via the keyboard 7 or the mobile telephone 15 or a different manner of input. For example, the patient data may be obtained from the patients file in the hospital management system, his personal health record or his mobile device 15, which comprises or is itself linked to sensors and/or includes an application enabling to input or provide such data. Such an application may also capture data from or provided via speech input by the user. This may be supplemented by data provided by a medical practitioner or data obtained from diagnostics equipment such as an imaging system 9.

In step 22, medical intervention data is obtained for the medical intervention to be evaluated. Again this data may be provided by the medical practitioner, or a desired medical intervention to be evaluated may already be registered in the hospital management system in the patient’s file. In embodiments wherein multiple medical interventions are to be compared with each other, step 22 may comprise the obtaining of medical intervention data for each of these medical interventions.

Next, in step 24, based on the medical intervention data obtained in step 22, the decision support system 1 will obtain a prediction model data from database 10. The prediction model data will relate to the medical intervention or medical interventions to be evaluated. In principle, applying the prediction model data for the prediction model obtained for database 10 to the characteristics provided in the medical intervention data and the patient data will provide a general prediction of the outcome of a certain medical intervention which is to some extent already individualized based on the variables available in the prediction model data. For example, if the clinical study underlying the prediction model discriminates between patients of a certain age, then age is a variable and if the patient to be reviewed is of a certain age, say 60 years old, the prediction model may provide the general probability values for a 60-year-old patient.

In accordance with the present invention, in step 26 the patient characteristics will be analyzed to identify certain specifics of the patient or his/her disorder that may give rise to a modification of the prediction model obtained from database 10 in step 24. For example, the specific patient may suffer from a comorbidity that significantly impacts the probability of a positive outcome of the medical intervention, or which may increase the probability of developing a certain side effect. In step 26, from the database 10, modifier data is obtained that relates to the impact of the occurrence of this specific patient characteristic on the medical intervention results. For example, the modifier data may indicate to which extent the presence of a certain biomarker may give rise to the developing of a specific side effect such as chronic headaches. The modifier data obtained in step 26 is used in step 28 to modify the prediction model identified in step 24. The modified prediction model obtained in step 28 by modification will then be used in step 30 to determine the individualized predicted outcome of the medical intervention evaluated. The evaluation result of step 30 may be provided back to the user via his mobile telephone 15, or for example to the medical practitioner via his computer 32.

Alternatively and advantageously, another possibility is that a speech processing unit (which may be a software model) translates the predicted outcome 30 into spoken language. The spoken language may for example be presented to the user via a speaker 33. This may be the internal speaker of the telephone 15. The inventors have gained the insight that the information is conveyed much more effective via spoken word then via text on screen. Therefore, the outputting of the predicted medical intervention results from step 30 via speaker 33 will be advantageous to the patient. A further specific example of a method in accordance with the present invention is provided in figure 3. Figure 3 provides an example of a comparison of two different medical interventions: medical intervention A and medical intervention B. In step 20, patient characteristics 35 are provided to the decision support system. These patient characteristics 35 may for example include image data 351, specific tumor data about a tumor to be treated (e.g. the type of tumor), clinical data 353 and tumor biology or genetics 354. The patient data 35 may further include specifics of the patient itself, such as gender, age, biometric information, comorbidity, habits, profession, patient preference. With respect to patient preference, it is possible that the patient has indicated that in view of his profession or because of any other reason, a certain preferred outcome is desired. For example, the patient may have indicated that he or she does not want to lose a certain body function because this body function is important for him or her to perform his profession.

Next, for each medical intervention identified, a prediction model 38 will be obtained from the database 10. For example, for medical intervention A prediction model 381 will be identified in the database. For medical intervention B, prediction model 383 will be obtained from database 10. Next, based on the patient data the prediction models 38 will be modified using modifier data. In figure 3, for medical intervention A the modifier data will be a further prediction model 382 which more accurately predicts the outcome of the medical intervention for patients with the give patient characteristics 35. Furthermore, for medical intervention B, because the patient has been associated with the occurrence of a specific biomarker, the modifier data consists of data from a case study 391 to the impact of the biomarker on medical intervention B. As may be appreciated, the modifier data may relate to any number of patient characteristics identified in the patient and is not limited to only one specific characteristic or one biomarker.

The combined prediction model which is corrected using the modifier data, is used to predict an outcome for each of the medical interventions, medical intervention A 401 and medical intervention B 402. The predicted outcome is generally referred to as 40. By comparison, in step 42, the patient may be presented a review of each of the medical interventions for his specific case. This may allow the patient to choose a preferred medical intervention or may enable the medical practitioner to advise the patient on one of the medical interventions for his case. Optionally, cost information may be added that enable to perform a financial evaluation of each of the medical interventions as well. For example, if medical intervention A would be preferred but the costs of medical intervention A are such that this will severely impact the financial situation of the patient after curing the disease, the patient may prefer that alternative medical intervention B instead. The cost comparison will be performed in step 45 and provided as an outcome in step 46. The cost comparison is only optional to the present invention.

Figure 4 schematically illustrates the performance of a clinical trial of a small group of patients. Typically, if a clinical trial is to be performed for certain patients suffering from a specific comorbidity which does not occur very frequently, the size of the population is too small to be able to perform a clinical trial reliably, and for that reason such a clinical trial cannot be carried out. In the situation of figure 4, the population of patients 50 contains two patients 501 and 502 that both suffer from the same comorbidity. Preferably a clinical trial would provide insight in the impact of this comorbidity on the medical intervention by medication 52, however, the size of the group (only two patients) is too small to perform such a clinical trial 55. In this case, no clinical trial will be carried out, and the patients 501 and 502 may be treated anyway without being completely certain of the effects or outcome, or the medical intervention will not be carried out in view of the uncertainty.

As illustrated in figure 5, the decision support system 1 of the present invention enables to perform an in silico trial to gain insight in the outcome of the medical intervention for the patients 501 and 502. An in silico trial is a clinical trial performed based on simulation only, reusing real data of patient treated in the past or using synthetic data with or without validated models. The difficulty with performing an in silico trial in reality is that the reliability of the outcome of such an in silico trial in principle is unknown. As illustrated in figure 5, a cohort 60 of patients 50 may undergo a certain clinical trial for the medical intervention with medication 52. The clinical trial 55 will provide a clinical trial result 58. The clinical trial 55 is a regular real clinical trial performed amongst the cohort of patients 60. The clinical trial results will be published and will be available in the database 10. The decision support system 1 may generate an alternative cohort 61 of virtual patients 51. The virtual patients 51 are not real patients, but are data elements providing a collection of patient characteristics generated based on statistics. Alternatively here one may also reuse real data of patients treated in the past. The virtual cohort 61 will be used to perform an in silico test trial. For the medication 52, a prediction model 381 will be identified in the database 10. For each of the patients 51 of the virtual cohort 61, modifier data 382 will be obtained from the database 10 and will be used to perform the in silico test trial 65. The trial results of the test trial will be available in step 68 as test results. Then, the decision support system 1 performs a comparison step 70 between the test results 68 of the in silico test trial 65 and the trial results 58 of the real clinical trial

55. This comparison is performed in step 70. If the results of the clinical trial 55 and the test results 68 are similar and do not deviate too much from each other, this is an indication that the in silico test trial is a success, the calibration and validation succeed and the in silico trial could be reused to test a new hypothesis or to make a prediction for an individual patients. If the test results 68 differ too much from the clinical trial results 58, this is a clear indication that the in silico test trial may have to be further modified. In step 72, the in silico test trial may be further modified by identifying additional modifier data to the prediction models 81. Alternatively, if this cannot be performed successfully it may also be decided to simply cease the in silico trial here. However, if the in silico test trial was a success, then in step 77 the in silico trial will be performed on a further cohort 73 of patients 74, which include patients 741 and 742 suffering from a same comorbidity. In principle, the number of participants suffering in the cohort 73 and the number of participants from the comorbidity may be varied based on the trial to be performed. The in silico trial 77 will then provide an in silico trial result 78. This can be used as modifier data or as prediction model in the decision support system.

The above may be applied advantageously in the preparation of clinical trials or to evaluate whether or under what conditions a clinical trial may take place to test a certain hypothesis. For example, in order to calculate a sample size needed for a real world clinical trial (e.g. how many patients are needed to confirm an difference x between arm 1 and arm 2 for a given alpha and beta), the invention according to these embodiments allows to test different hypothesis in in silico clinical trials, on real data or synthetic data before a real trial is performed. Also, in order to stimulate inclusion in clinical trials, the invention according to these embodiments can be used for “trial Patients Decisions Aids” (tPDA) with models or the trial assumptions used to calculate sample size or the models (form preclinical data, or calculating the probability of cure and toxicity). Furthermore figure 6 schematically illustrates how prediction model data from a clinical study 80 may be obtained in the decision support system 1. In step 82, a nomogram 80 of the clinical study may be analyzed and digitized. Next in step 84, the digitized values are analyzed to determine corresponding probabilities. For example a logistics regression model may be applied or a cohort 85 of virtual patient may be generated by the decision support system. In step 86, from the results of step 84, the coefficients are derived which describe the model, for example by fitting. For example, for each of the patients in cohort 85, the result of the prediction model is obtained from the nomogram data obtained in step 82. This will produce prediction model data 89 which is provided as data to the database 10. Thereafter, the prediction model data 89 may be analyzed by a (linear) regression model to obtain coefficients that describe the prediction model for any arbitrary case. The results may be included in the database 10 to replace or be added to the prediction model data 89. A detailed description of such a method will be described further below in this document for a particular example.

Example case:

Evaluating medical intervention types for the medical intervention of prostate cancer.

Abstract

Studies have found that survival after radical prostatectomy (RP) and external beam radiotherapy (EBRT) as primary medical intervention for prostate cancer (PCa) is similar, but the risks of toxicities are not. The aim of this study was to build a decision support system (DSS) using a decision-analytic model (individual state-transition model) based on predictive models (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD)) for estimating tumor control and toxicity probabilities for both RP and EBRT for low to intermediate risk localized PCa. We performed analyses on a synthetically generated dataset of 1000 patients with realistic clinical parameters, externally validated our approach by comparing our results with those from randomized clinical trials, and subsequently set up an in silico clinical trial for elderly patients . We assessed the cost-effectiveness (CE) of the DSS for medical intervention selection by comparing it to randomized medical intervention allotment as a proxy for current clinical practice. Using the DSS to decide upon medical intervention (RP or EBRT) for the synthetic dataset , 55% were selected for RP and 45% for EBRT. Upon comparison to published results, differences with the simulations of late toxicity and biochemical failure never exceeded 2%. The in silico trial showed that for elderly patients, toxicity has more influence on the decision than TCP, and the predicted QoL depends on the initial erectile function. The DSS is estimated to result in cost savings (€1005, 95% CI: €159- €1822) and more quality adjusted life years (QALYs; 0 17 years, CI: 0 €0-0 33) than randomized medical intervention selection.

This modelling study shows that DSS based medical intervention decisions will result in a clinically relevant increase in the patients’ quality of life and could be used for in silico trials.

Introduction

Prostate cancer (PCa) is the second most commonly diagnosed cancer for men, accounting for 13 5% of new cancers diagnosed in 2018 of which 40% are low to intermediate risk localized PCa in the Netherlands. PCa is a topic of heightened research interest, with new biomarkers and medical intervention modalities being tested at a high rate. The leading choices for managing clinically localized PCa are external beam radiotherapy (EBRT), radical prostatectomy (RP), brachytherapy, and active surveillance For low and intermediate risk PCa, active surveillance is often proposed (~70% and ~30% respectively), and of the active medical intervention options, RP and EBRT are recommended most often (~50% and ~45%), according to the Netherlands Cancer Registry. However, no consensus has been reached as to which is superior in terms of effectiveness and/ or toxicity, both due to the varying spectrum of toxicities as well as the difference in incidence. The medical intervention decision is often based on doctor preference, and to a much lesser extent on patient preferences and patient-specific characteristics, or expected outcome such as (biochemical recurrence-free) survival or toxicity.

Currently, there are no studies available that help assess the individual benefits of EBRT versus RP based on patient characteristics, even though it is suspected that parameters such as age, BMI, tumor grade, and pre-medical intervention prostate specific antigen (PSA) levels do influence both recurrence-free survival and toxicity. The importance of personalized medicine has become progressively evident, and medical intervention selection for PCa is no exception. An important step towards personalized PCa medical intervention would be a clinical decision support system (DSS) that aids in the decision between RP and EBRT. In addition, the possibility of very cost-effective in silico trials (individualized computer simulations used in the development of drugs, devices or interventions) promise to improve clinical research through better design, more transparent and detailed information about possible results and greater explanatory power in interpreting side effects, as well facilitating exploration of interactions with the individuals’ biology and of the long-term or rare effects.

Our hypothesis was that the DSS can be used for in silico trials and can accurately replicate the results from published studies. We also hypothesized that the use of a DSS for medical intervention selection results in better tumor control, less toxicity, increased patient quality of life (QoL), and improved cost-benefit ratio when compared with current clinical practice based on tumor boards or medical specialist opinion.

The aim of this study was to build such a DSS using predictive models for estimating tumor control and toxicity probabilities for both RP and EBRT for low to intermediate risk localized PCa patients and validate this by comparing to published clinical trials. We also set up an in silica trial using this model based approach to assess outcome for elderly patients. Additionally, we compared the cost-effectiveness (CE) of applying this DSS to random medical intervention decisions as a proxy for current clinical practice.

Materials and, Methods

DECISION SUPPORT SYSTEM

Markov model

The target population consisted of overall tumor stage T1-T2 PCa patients who were eligible for active medical intervention (i.e. EBRT and RP). The DSS was developed by constructing an individual state-transition model to estimate the effects and associated costs of medical intervention with RP vs. EBRT for each patient. Based on patient-specific parameters (e.g. age), and medical intervention type (EBRT or RP), probabilities to develop long-term toxicities including rectal bleeding, urinary incontinence, and impotence, or a combination, are calculated. After medical intervention, patients have a risk of progressing to the recurrence state, which is dependent on patient specific parameters (e.g. Gleason score), after which they can develop metastatic disease and subsequently progress to PCa-related death.

Furthermore, from any health state, it is possible to die of causes unrelated to cancer (see figure 7). Figure 7 illustrates a summary of the Markov model 100. Ovals 101 represent different health states, arrows 102 represent transitions between health states. Dashed arrow lines 103 are for intelligibility purposes. Patients start in the disease-free state, with either all toxicities, ED and UI or UI only, and as time passes, they can recover from toxicity, or progress into the biochemical progression state. Death unrelated to cancer can occur from any health state, cancer related death only from the metastatic disease state.

The DSS then provides a comparison of the tumor control probability (TCP), probability of chronic erectile dysfunction (ED), chronic urinary incontinence (UI), and late rectal bleeding (RB), as well as a comparison of expected costs and quality adjusted life years (QALYs).

Several assumptions are made in this model, the most relevant four being: i) All patients start with ED and all RP patient start with UI (at the first cycle), and a percentage of EBRT patients start with RB; ii) When a patient has developed biochemical failure, it is assumed that toxicity does not affect the health related quality of life (QoL) or the costs anymore (i.e. QoL and costs are only driven by biochemical failure and not by toxicity), so toxicity is only incorporated in the disease free health state; iii) We assume that after progression, no secondary medical intervention takes place and the costs of this health state are the same as for disease free; iv) The transition probability from recurrence to metastatic disease is the same for all patients for both medical interventions.

Moreover, a cycle time of one month and a time horizon of 20 years are used. This was chosen since the survival of low-intermediate prostate cancer is high while increasing the time horizon beyond 20 years would unnecessarily increase model uncertainty. Utility values

In order to quantify the relative importance of various health outcomes using a common measurement unit, utility is used as a metric to assign weights to health states on a scale ranging from 0 (for dead) to 10 (for perfect health). Health-state specific utilities, also incorporating medical intervention related toxicities (and all possible combinations), were retrieved from literature (see table 1). QALYs are obtained by multiplying these utility values by the time spent in the corresponding health-state. Table 1 provides a list of utility values used in the state transition model. In order to account for the baseline utility for men living in the Netherlands, we used a model that calculated the age dependent health related quality of life (HRQoL) for different countries and applied it as a multiplicative factor to the health- state specific utilities.

Predictive models

The transition probabilities were estimated per individual in order to make this DSS patient specific and ready for precision medicine applications. The individual probabilities of progression after medical intervention, and the risk of developing toxicities, were calculated using a selection of regression models or nomograms from the published literature (see table 2 below), adherent to the TRIPOD statement. For nomograms, the coefficients or intercepts were derived (if not reported) by reading the nomogram and using interpolation and fitting. Table 2 provides an overview of the literature models used for the state transition probabilities. Rectal bleeding does not typically occur after RP, so the transition was set to zero for this medical intervention type. cores [n] Negative biopsy cores [n] Schaake 1b EBRT Mean Trigone dose Late UI after AUC = 0.66 et (243) [Gy] 3 years al. 2018²² Matsushit 2a RP Age [years] Recovery from AUC = 0.71 a et al. (2849) BMI [kg/m²] UI after 1 2015^{* 23} ASA score [I/II / year III/IV] Urethral length [mm] Alemozaff 3 Both Age [years] Erection AUC = 0.77 ar et al. (524 + 241) Nerve-sparing [y/n] recovery for RP, 2011²⁴ PSA [ng/ml] after 2 years AUC = 0.87 ADT [y/n] for EBRT Valdagni 2a EBRT Anticoagulants [y/n] Late RB after C-index = et al. (1132) Diabetes [y/n] 3 years 0.62 2012^25,26 Hemorrhoids [y/n] Irradiation PN [y/n] Hormonal therapy [y/n] Abdominal surgery [y/n] Mean rectal dose [Gy] V75 rectum [%] TRIPOD: Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis; EBRT: external beam radiotherapy; RP: radical prostatectomy; n: number; ADT: androgen deprivation therapy; PSA: prostate specific antigen; BED: biologically effective dose; BFFS: biochemical failure-free survival; R2: coefficient of determination; Gy: gray; BMI: body mass index; C-index: concordance statistic; RFS: regression free survival; UI: urinary incontinence; ASA: American Society of Anesthesiologists; PN: pelvic nodes; V75: volume receiving at least 75 Gy; RB: rectal bleeding *Due to incomplete information, some model coefficients had to be derived from the nomogram. **Recurrence was defined as biochemical failure, initiation of secondary therapy, distant metastases or prostate cancer related death. Table 2 Validation Synthetic dataset Since no complete patient cohort was identified for which all required input parameters are known, we generated a synthetic patient dataset. A patient is described by a set of parameters, which we randomly assigned by drawing values from a distribution. We chose the distributions based on the patient cohorts on which the models had been developed. We assumed that the clinical parameters were independent from one another and generated a synthetic patient dataset including 1000 patients, so that the mean and associated error of the generated clinical parameters matched that of the original datasets. Validation of NTCP and TCP models In order to validate the combination of our NTCP and TCP models, we compared the simulation results to the published results. The reliability of the model, and the measure in which the synthetic dataset reflects real patient datasets was assessed by generating the synthetic patients so that they match the reported clinical parameters from actual trials, such as age, Gleason score, PSA values, and T-stage. Non-reported clinical parameters were kept the same as the original synthetic dataset. The relevant outcomes, such as biochemical free survival or toxicity, were then compared between the simulation and the clinical trial. For the EBRT |/} ratio (a measure of the fractionation sensitivity of the tissue) of PCa, we used a value of 15 Gy. In order to assess the credibility of the predicted biochemical free survival, the model was compared to a retrospective study that compares EBRT to RP, and to assess the effect of hypofractionation we compared the biochemical free survival to the CHHiP trial. In order to validate the models predictions of toxicity, we compared the model results to a literature study on patient reported outcomes after EBRT or RT.

In silico trial

A DSS such as this, in combination with the synthetic patient dataset, could function as an in silico clinical trial, a precursor to actual clinical trials, in order to improve study design or explanatory power. We demonstrated this by generating a patient dataset with patients aged 75-90 to test the outcomes of different medical interventions for elderly patients, an often underrepresented group in clinical trials. We adjusted the pre-medical intervention erectile function to be an average of 15%, to reduce the impact of ED on the outcome.

Results

Synthetic data characteristics

When applying the DSS on a synthetically generated patient dataset of 1000 patients with clinical parameters similar to those on which the predictive models were built, 55% of the patients had a higher number of QALYs for RP, and 45% for EBRT. The patients for whom EBRT was chosen had a higher mean age (64 versus 59 years), higher mean prostate specific antigen (PSA) values (8.3 versus 6.8 ng/ml), and a higher percentage of T2 stage (49% versus 24%), as shown in table 3. The percentage of patients in each health state over time is shown in figure 8. Figure 8 depicts distributions 111 and 112 of the patient population over different health states over time. The solid lines in figure 8 represent the radiotherapy arm, and the dashed lines represent the prostatectomy arm.

Parameter name EBRT mean RP mean (SD) P

(SD)

Age (years) 63.8 (10.7) 58.8 (9.1) «0.001

PSA (ng/mL) 8.3 (3.5) 6.8 (3.5) «0.001

Validation

We compared the outcome reported in three separate papers against the same outcome simulated by our DSS on a synthetic patient dataset. The results are reported in table 4 below. The progression free survival reported in Aizer 2009 was very similar to our simulation results. Dearnaley 2016 published the results of the CHHiP trial and compared the results of different fractionation plans. When comparing these results to what was simulated by our DSS, the simulation was consistently around 7% lower, so the effect of hypofractionation was simulated well, but not the absolute outcome. Donovan 2016 was used for toxicity comparisons, and showed that the DSS simulated late toxicity best, but was less accurate for acute toxicity. The relative differences between the simulations and the studies for the different medical intervention modalities were similar, and the conclusions coincided. The most notable discrepancy is for acute ED and acute RB, which are overestimated by the simulations. For ED this can be explained as in the model, acute ED is based on assumptions, not models. In silica trial

We performed an in silica trial on the synthetic dataset by increasing the age of all patients to be between 75 and 90, but leaving all other clinical parameters as they were. We found that TCP for RP was marginally higher than for EBRT (HR: 0 98). The risk of chronic UI is much higher for RP (HR: 1086), which resulted in higher costs. However, the EBRT did result in more QALYs than RP. The DSS selected EBRT for 85% of the patients. We repeated this analysis while assuming that 85% of the patients had ED before the start of medical intervention, with RP gaining more QALYs, and the DSS selecting RP over EBRT in 75% of the patients, see table 5 below. Discussion

In this example case study we described a clinical DSS for the medical intervention of PCa patients with either EBRT or RP, and tested this on a synthetic patient dataset. We validated the DSS against published clinical studies and set up an in silica trial for patients between 75 and 90. We also assessed the CE of a medical intervention allotment strategy based on the DSS compared to a randomized medical intervention allotment strategy. Our first hypothesis was that we can accurately replicate results from published studies, which we aimed to confirm generating synthetic datasets with clinical parameters similar to published trials. The DSS largely replicated the published results accurately. The relative differences between the medical intervention modalities and fractionation plans was replicated by the model and the conclusions of the DSS and the studies were in agreement. We also performed an in silica trial exclusively including elderly patients with or without ED using the DSS and found that for the first group, EBRT had the preference, and for the second, RP performed better in terms of QALYs. Additionally, we observed that a medical intervention selection strategy based on the DSS would improve tumor control and reduce toxicity as opposed to randomized medical intervention selection.

These results imply that, when deciding between RP and EBRT for a given patient, making the right choice can improve overall QoL, and that this decision should not be random. The DSS offers the possibility to combine a large quantity of clinical parameters, predict NTCP and TCP and quantify these risks into a single metric for different medical intervention options. This has the potential to improve the decision-making process, along with other factors, such as incorporating patient preferences. The development of a DSS fits well into the current trend that strives for personalized medicine, and the results presented in this study confirm the added benefit of such tools. The application of the DSS for in silica trials has great potential benefits, not only by improving the design of clinical trials through precursory simulations, it also has the benefit of being able to apply different medical interventions to the same “patient”, which allows for a more objective comparison. Another advantage of the DSS is that it is detailed and can further extended with other disease management options such as brachytherapy or active surveillance. It can also be used as the basis for an individualized patient decision aid (iPDA).

This particular example case study had several limitations. The first one is that this is a model -based study, using models which were trained and validated on different cohorts. The models were selected based on how recently they were published, the number of patients included, and whether or not they used clinical parameters and TRIPOD level. We also attempted to make sure we only selected models trained on patients with similar medical intervention modalities and similar clinical parameters, however not all clinical parameters were reported. Also, the correlation between clinical parameters is not reported, and when generating the synthetic dataset, no correlations were assumed. Different outcomes of the models were validated on different studies, so the DSS as a whole has not been validated on a single patient population. However, the acquisition of a dataset where not only all the clinical parameters are reported, but also long term follow-up data for TCP and toxicity for both medical intervention arms might not be feasible. Also, the NTCP models used doctor-reported outcomes as endpoints, while the validation was done on patient reported outcomes. The NTCP model for bowel toxicity becomes very inaccurate at probabilities below 5%, and many newer studies find very low incidence rates of grade 2 bowel toxicity. This means that the bowel toxicity predictions have limited value in its current form and should be retrained on more recent patient datasets. In its current state, the DSS does not take into account patient preferences, but uses average utility values obtained from a population. However, risks of different types of toxicities are what often drive medical intervention decision making, and patient preferences should be taken into account.

In conclusion, this study is exemplary of a detailed, personalized medical intervention DSS that aids in the choice between EBRT and RP for low to intermediate risk PCa patients. This DSS could be used for in silico clinical trials when applied to a synthetic dataset, which would be a valuable precursor to clinical trials. The results suggest that the full development and clinical application of this DSS would improve the quality of patient care and would be an important step towards personalized and participative medical intervention decisions.

Example case 2:

Evaluating the optimal vaccination types for a specific patient in the context of a medical intervention to prevent infectious disease

The whole world has been hit in 2020 by the COVID-19 pandemic caused by the novel coronavirus Severe Acute Respiratory Syndrome Coronavirus 2 (SARSCoV-2). There are more than 190 vaccines in development (of which none of them has a 100% coverage) and the virus is continuously mutating. The question will arise which vaccin should a given person use to have the best and the longest protection and the lowest rate of side effects? The objective of this analysis is to create an algorithm able to identify the vaccine that will give the best coverage (at present time and in the future, taking into account the HLA haplotype of the patient and the predicted mutations of the virus) with a sufficiently broad repertoire of T-cell epitopes with the lowest likelihood of side effects in particular anaphylactic shock, fatigue and off-target autoimmune response. The SARS-CoV-2 proteome has been sequenced across the most frequent HLA-A, HLA-B and HLA-DR alleles in the human population, using host- infected cell surface antigen presentation and immunogenicity predictors from the NEC Immune Profiler suite of tools. This allowed to generate comprehensive epitope maps. This epitope map has then been used as input for a Monte Carlo simulation to identify statistically significant “epitope hotspot” regions in the virus that are most likely to be immunogenic across a broad spectrum of HLA types. Critically epitope hotspots that shared significant homology with proteins in the human proteome have been identified to reduce the chance of inducing off-target autoimmune responses. This exercise can be redone for a specific given patient taking into account his HLA type and his proteome (facultative). The antigen presentation and immunogenic landscape of all the nonsynonymous mutations across 5.000+ different sequences of the virus, has been analyzed to identify a trend whereby SARS-COV-2 mutations are predicted to have reduced potential to be presented by host infected cells, and consequently detected by the host immune system. Then the epitope hotspots that occurred in less- conserved regions of the viral proteome are removed. This exercise has to be repeated regularly based on the new real information related to the real mutations rather than the predicted one. A database of the HLA haplotypes of approximately 15,000 individuals has been built to develop a In Silico “digital twin” type simulation to model how effective different combinations of hotspots would work in a diverse human population. A Monte Carlo and digital twin simulation, can then be done for a specific patient using HLA type of the patient, the last data on SARS-CoV-2 proteome, the mutation trend and the antigen presentation to the infected-host cell surface, immunogenicity predictions of the NEC Immune Profiler. Finally the preference of the patients and his “risk profile” can be taken into account to choose the vaccine that will give the best coverage/side effects ratio. Embodiment description:

Getting model coefficients from a Nomogram

Many prediction models considering outcome in cancer medical intervention are based on regression models. In order to improve readability and interpretability by medical specialists, these models are often converted to nomograms. However, when the need arises to apply the model in a decision support system in accordance with an embodiment of the invention, alone or in combination with other models, it is necessary to extract the original model coefficients. These, however, are not always published or made available by the authors. To this end, the invention in accordance with these embodiments provides a simple method to extract the coefficients from nomograms for regression models.

Reading the nomogram (step 82)

Nomograms typically need to processed manually by a user to obtain the information therefrom. In order to simplify this, the nomogram in accordance with these embodiments may be automatically read such as to enable using the prediction model data in the decision support system. For visual comfort, in this embodiment a nomogram was converted to a spreadsheet. The nomogram 120 used for this demonstration wat taken from {Valdagni 2008}, and shown in Figure 9. To automate the reading of the nomogram 120 of figure 9, table 6 below was extracted by carefully reading the points on the nomogram 120. Since the relationship between the parameters and the points on the nomogram are binary or linear, this information is enough to automatically count the number of points for any patient using the following equation: Note that for and the relationship with the parameters is inverted, as these have a negative impact on the risk of toxicity. Converting points to probability A logistic regression model can be described as follows: A similar relationship is expected between and the Probability of G2-G3 acute toxicity . In order to obtain the values for intercept and coefficients , we need to, as accurately as possible, read the total points related to the different values for on the nomogram, as shown in Table 10. In order to get a better form for the logistic regression model, we need to normalize for the points assigned for the use of anticoagulants and for the use of hormones, so the total number of points is calculated as follows: The total points row in table 10 may be adjusted accordingly, and the logit of the probability should be calculated, as shown in Table 11. Obtaining the regression model A linear line may now be fitted to estimate the relationship between P_total and logit(p), as seen in the graph 130 in Figure 10. We now know the coefficients and the intercept, which are shown in Table 4. Table 4 Vahd h]dlh i]Z ejWa^h]ZY dYYh gVi^dwh) hd lZ XVc hZZ i]Z eZg[dgbVcXZ d[ i]^h bdYZa+ Example modifier data: How to add the effect of a gene polymorphism or a tumour mutation to an existing model? The community has been working for more than 60 years on dose response curves or predictive models using conventional input (dose, TNM). The world of genetics, focusing one gene polymorphism or SNP of normal cells (e.g. DNA of lymphocytes or saliva cells) or tumour mutations is rather new. The question we asked is: can we combine the two approaches without having to develop new models from scratch? The answer is yes, as explained below. Here we illustrate a case where the effect of validated SNP is added to another validated modes of dose response curve. Using the minor allele frequency (MAF) we can estimate the prevalence of the SNP in the population. Using the odds ratio (OR) we can split a dose response curve into two curves: one for the population with SNP, one for the population without. The two SNPs most associated with rectal bleeding, that approach meta- analysis significance, were selected for the initial proof-of-concept: rs141044160 (SNP 1) with an OR of 2.68 (P = 2.26 × 10^y3) and rs7432328 (SNP 2), with an OR of 3.36 (P = 3.32 × 10^y3). N]Z MHJwh lZgZ ^cXdgedgViZY ^cid i]Z YdhZ gZhedchZ XjgkZ jh^c\ V published mathematical method [1] described below. The probability of late rectal bleeding is described by the following equation: Where: Here is the equivalent uniform dose, determines the steepness of the response, and is the amount of dose given for the probability of late rectal bleeding to be 0.5. We wish to separate the population into two groups (low risk group) and (high risk group) with complication probabilities and for the corresponding groups. The prevalence factor is given by: and is assumed to be independent of dose. The prevalence can be calculated using the MAF. To determine the equations for Po and Pi, we must find the corresponding parameters and ^mo and and ®i. To find DQ we must find the value for EUD where Po is 0.5.

We can do this using equation (4) and filling in Po — 0-5;

Using equation (1) and (2) we can obtain Do. First we must write as a function of P: ) Filling in P = P results in the following equation for

The present invention has been described in terms of some specific embodiments thereof. The method and decision support system of the present application may be applied as a method for:

- medical intervention decision for a medical practitioner or computer using a DSS build from multiple (externally) validated predictive models integrating different type of data (clinical, imaging, biological genetic, environmental) and testing different medical intervention type;

- facilitate participative medicine, this method and decision support system can be used for individualized Patients Decisions Aids (iPDA) with models calculating the probability of cure, toxicity, financial toxicity;

- facilitate inclusion in clinical trials, this method and decision support system can be used for trials Patients Decisions Aids (tPDA) with models coming form systems biology y, mechanistic modeling, animal or patient data, calculating the probability of cure and toxicity;

- to calculate a sample size needed for a real world clinical trial; and

- to test hypothesis knowing that this trial will not be feasible (e.g. too few patients, no payer, ethically unacceptable) by performing an In Silico clinical trials, using real world data and/or synthetic data, based on this method and decision support system.

It will be appreciated that the embodiments shown in the drawings and described herein are intended for illustrated purposes only and are not by any manner or means intended to be restrictive on the invention. The context of the invention discussed here is merely restricted by the scope of the appended claims.

Claims

Claims

1. Method of operating a decision support system for evaluating a medical intervention plan for the medical intervention of a disorder in a human or animal body, the method comprising the steps of: obtaining, by the decision support system, patient data indicative of one or more patient characteristics for an individual patient; obtaining, by the decision support system, medical intervention data indicative of the medical intervention plan for the medical intervention; identifying, based on the medical intervention data, at least one prediction model associated with the medical intervention plan; applying, based on the patient data, at least one modifier model for modifying the at least one prediction model based on the patient characteristics of the individual patient, such as to obtain a modified prediction model; and determining, based on the patient data and using the modified prediction model, a medical intervention outcome in terms of one or more probability values associated with one or more health status events indicative of medical intervention result probabilities.

2. Method according to claim 1, further comprising a step of obtaining preference data indicative of a health status preference or risk profile preference of the patient in relation to the medical intervention associated with the disorder; and comparing the preference data with the one or more probability values for determining an agreement level therebetween.

3. Method according to any one or more of the preceding claims, wherein the prediction model provides statistics data, wherein the statistics data enable to determine, based on at least one of the patient characteristics, at least one of the probability values associated with at least one of the health status events.

4. Method according to claim 3, wherein the at least one prediction model is a regression model providing a predictable dependency between the at least one of the patient characteristics and the at least one of the probability values, wherein the method comprises a step of: receiving, by the decision support system, a nomogram wherein the nomogram quantitatively describes the dependency between the at least one of the patient characteristics and the at least one of the probability values; and analyzing the nomogram such as to determine, from the nomogram, coefficients of the dependency.

5. Method according to claim 3 or 4, wherein the modifier model includes modifier data, the modifier data being indicative of an impact of a particular patient characteristic state on the at least one of the probability values associated with the at least one of the health status events.

6. Method according to claim 5, wherein the modifier data includes a hazard ratio associated with the particular patient characteristic state, such as a hazard ratio associated with a biomarker.

7. Method according to any one or more of claims 3-6, wherein the modifier model includes one or more further prediction models different from the at least one prediction model associated with the at least one medical intervention plan, the further prediction models enabling to modify the at least one prediction model such as by averaging the probability values or correcting for occurrence of a particular patient characteristic state.

8. Method according to any one or more of the preceding claims, wherein the patient characteristics include one or more elements of a group comprising: demographic variables, such as gender, age; one or more body parameters, such as body size, weight, body mass index, general status or fat percentage; genetic characteristics, such as gene polymorphism, human leukocyte antigen, haplotype or somatic or germ-cell mutations; co-morbidities; one or more body- or physiological characteristics, such as hair color, skin type, speed of speech, type of speech, number of steps done per day, blood type; biological age; imaging data; data acquired from sensors; user input; medical status data or diagnostic data, such as medications taken, illness progression status, allergic response data; medical history data, such as medication history, earlier received medical interventions; behavioral data, such as practiced sports, physical exercise frequency, smoking habits; location data; climate or weather data, such as temperature, humidity, precipitation, pollution, allergens, air particles; environmental or social data, such as risk of infection, occurrence of an epidemy or pandemic, risk of terrorist attacks.

9. Method according to any one or more of the preceding claims, wherein the medical intervention data include one or more elements of a group comprising: indication of a medical intervention type, such as medical intervention by surgery, interventional radiology, drug therapy, irradiation therapy or radiotherapy, tumour- treating fields, movement therapy; medical intervention specifics, such as medicine data of a particular type of medicine, a data of a particular type of surgery, data relating to enrichment of the microbiome, data relating to treatment with a genetically modified organism (virus or bacteria), data relating to treatment with stem cells, data relating to geroprotectors.

10. Method according to any one or more of the preceding claims, wherein the health status events include one or more elements of a group comprising: cured; recurrence of the disorder after a predetermined time; decreased or accelerated progression of the disorder; occurrence of post-medical intervention adverse effects, such as a medical intervention related disorder or complication, medical intervention related damage to the body, loss of body function; quality of life, depression, happiness, death.

11. Method according to any one or more of the preceding claims, wherein the step of determining a medical intervention outcome determining a health state score based on the probability values associated with the one or more health status events.

12. Method according to any one or more of the preceding claims, wherein the method comprises comparing a first medical intervention with a second medical intervention for the medical intervention of the disorder in the human or animal body, wherein the step of obtaining the medical intervention data comprises: obtaining first medical intervention data indicative of the medical intervention plan for the first medical intervention, and obtaining second medical intervention data indicative of the medical intervention plan for the second medical intervention; wherein the step of identifying the at least one prediction model comprises: identifying at least one first prediction model for the first medical intervention and identifying at least one second prediction model for the second medical intervention; wherein the step of applying the at least one modifier model is performed for at least one of the first or second prediction model; and the step of determining the medical intervention outcome comprises determining a first medical intervention outcome based on the first and the second prediction model, as modified by the at least one modifier model, such as to yield a first and a second medical intervention outcome; wherein the step of comparing the first medical intervention with the second medical intervention is performed based on the first and a second medical intervention outcome.

13. Method according to any one or more of the preceding claims, wherein the step of obtaining, by the decision support system, patient data comprises: generating, by a controller, patient data for a plurality of individual virtual patients, wherein each individual virtual patient includes one or more patient characteristics associated therewith, the plurality of individual virtual patients thereby forming a virtual cohort.

14. Method according to any one or more of the preceding claims, as far as dependent on claim 2, wherein the preference data is obtained by at least one of: the patient, or a prediction based on preference data obtained from one or more further patients for which the preference data has been captured, such as patients having same or similar patient characteristics.

15. Method according to any one or more of the preceding claims, as far as dependent on claim 2, wherein multiple times during a time period the method comprises performing the steps of: identifying the medical intervention model, applying the modifier model and determining the medical intervention outcome; and at least one of the steps of: obtaining patient data, obtaining preference data or obtaining medical intervention data, for enabling to capture changes in the data over time.

16. Method according to any one or more of the preceding claims, as far as dependent on claim 8, wherein at least one of: the patient data are automatically provided from a proprietary database, phone, sensor or public database; or the patient data include diagnostic data, wherein the diagnostic data includes one or more elements of a group comprising: indication of a certain examination type, such as MRI of a body part or organ, CT of a body part or organ, echography of a body part, optical coherence tomography of a body part (e.g. skin or eye), fundoscopy, blood sample, colonoscopy, genetic test, PET- CT, PET- MR, functional test, bacteriological test, serological test, evoked potentials, test of the microbiome; indication of a specifics diagnostic method performed, such as a specific genetic test, or a particular type of microbiome analysis.

17. Method according to any one or more of the preceding claims, further comprising a step of outputting the determined medical intervention outcome.

18. Method according to claim 17, wherein the output is provided by at least one of: output via a display screen, a mobile phone, a computer, or a speech processor such as via spoken language.

19. Decision support system for evaluating a medical intervention plan for the medical intervention of a disorder in a human or animal body, the system comprising a controller and at least one of a memory or a data communication unit configured for communicatively connecting with at least one data repository, the memory or data repository being configured for storing instructions which, when stored in the memory or the at least one data repository, enable the controller to perform the steps of: obtaining, by the decision support system, patient data indicative of one or more patient characteristics for an individual patient; obtaining, by the decision support system, medical intervention data indicative of the medical intervention plan for the medical intervention; identifying, based on the medical intervention data, at least one prediction model associated with the medical intervention plan; applying, based on the patient data, at least one modifier model for modifying the at least one prediction model based on the patient characteristics of the individual patient, such as to obtain a modified prediction model; and determining, based on the patient data and using the modified prediction model, a medical intervention outcome in terms of one or more probability values associated with one or more health status events indicative of medical intervention result probabilities.

20. Decision support system in accordance with claim 19, configured for comparing a first medical intervention with a second medical intervention for the medical intervention of the disorder in the human or animal body, wherein the instructions further enable the controller to: perform the step of obtaining the medical intervention data such that it comprises: obtaining first medical intervention data indicative of the medical intervention plan for the first medical intervention, and obtaining second medical intervention data indicative of the medical intervention plan for the second medical intervention; perform the step of identifying the at least one prediction model such that it comprises: identifying at least one first prediction model for the first medical intervention and identifying at least one second prediction model for the second medical intervention; perform the step of applying the at least one modifier model for at least one of the first or second prediction model; and perform the step of determining the medical intervention outcome such that it comprises determining a first medical intervention outcome based on the first and the second prediction model, as modified by the at least one modifier model, such as to yield a first and a second medical intervention outcome; wherein the step of comparing the first medical intervention with the second medical intervention is performed, by the controller, based on the first and a second medical intervention outcome.

21. Decision support system according to any one or more of claims 19-20, wherein the instructions enable the controller to perform the step of obtaining patient data such that it comprises: generating, by the controller, patient data for a plurality of individual virtual patients, wherein each individual virtual patient includes one or more patient characteristics associated therewith, the plurality of individual virtual patients thereby forming a virtual cohort.

22. Decision support system according to any one or more of claims 19-21, wherein the system is communicatively connected with the at least one data repository for obtaining at least one of: the patient data, the medical intervention data, the at least one prediction model, or the at least one modifier model.

23. Use of a decision support system according to any one or more of claims 19- 22 as trial patients decisions aids for evaluation of a potential participation of an individual patient in a clinical trial.