1288875 九、發明說明： 【發明所屬之技術領域】 本發明係關於一種長期多項生理訊號自動處理系統及 方法’尤指一種藉由以圖控式介面實現之長期多項生理訊 號分析方法舆系統，而適用於醫院、各式醫療院所或其他 類似之裝置者。 【先前技術】 背景： . 睡眠窒息症（sleep apnea)、睡眠困難（dyssomnia)舆 • 類睡症（parasomnia)等等為常見的睡眠異常疾病，其中睡 眠困難包括了失眠（insomnia)及過眠（hypersomnia)，這些 病症不但影響患者的健康舆日常生活作息，長期下來將引 發許多併發症狀，嚴重的甚至會導致死亡。在臨床醫學上， 多項生理訊號監測舆睡眠期（sleep stage)判別為這些睡眠 疾病診斷舆治療的重要根據之一，根據患者的病徵與醫療 上的需求，可藉由檢閱長時間紀錄之生理訊號，如雎電圈 (Electroencephalogram, EEG) 、心 電圈 (Electro-cardiogram，ECG)、眼電圓（electrooculogram， 泰 EOG)舆肌電圈（Eleetro-myogram，EMG)等等，來判斷病灶 並記錄於電聪或醬療儀器裡。一般而言，針對一位病患紀 錄24小時、16個軌道的滕電訊號，就必須耗费將近leS個 Gigabyte的硬碟空間，相當於使用傳統颱波儀器8500張颺 波記錄紙，所以檢閱這些冗長且非穩態之生理訊號便顯得 十分的耗時舆费力。 由於數位時代的來臨舆電子技術的進步，數位化的生 理訊號紀錄儀器已逐漸取代傳統類比式儀器，針對這些所 12888751288875 IX. Description of the Invention: [Technical Field] The present invention relates to a long-term multi-path physiological signal automatic processing system and method, and more particularly to a long-term physiological signal analysis method 舆 system realized by a graphically controlled interface, Applicable to hospitals, various medical institutions or other similar devices. [Prior Art] Background: Sleep apnea, sleep difficulties (dyssomnia), parasomnia, etc. are common sleep abnormalities, in which sleep difficulties include insomnia and oversleep ( Hypersomnia), these conditions not only affect the health of patients, daily life and rest, will lead to many complications in the long run, and even lead to death. In clinical medicine, multiple physiological signal monitoring sleep stage is one of the important basis for the diagnosis and treatment of these sleep diseases. According to the patient's symptoms and medical needs, the physiological signal can be recorded by long time recording. Such as Electroencephalogram (EGG), Electro-cardiogram (ECG), electrooculogram (EOG) Eleetro-myogram (EMG), etc., to determine the lesion and record In the electric or sauce treatment equipment. In general, for a patient to record a 24-hour, 16-track Teng signal, it is necessary to use nearly 1.8 gigabytes of hard disk space, which is equivalent to using the traditional Taiwan wave instrument 8500 Zhangbo recording paper, so review these The lengthy and unsteady physiological signals are very time consuming and laborious. Due to the advent of electronic technology in the digital age, digital biometric signal recording instruments have gradually replaced traditional analog instruments for these 1288875
II
I 紀錄的數位生理訊號，利用電觸進行分析舆處理，能快速 且有效率的分析出生理訊號所代表的特徵與意義，提供醫 療人員在診斷與判讀上之重要參考。然而，針對滕電訊號 或其他生理訊號，仍必須發展出更為快速且強健之數位訊 號處理、辨識演算法與針對長期監測生理訊號所設計之分 類系統，以得到更精確且有效率的分析結果，並節省大量 的人力資源舆時間，而這些研究工作亦為目前國内外研究 單位所努力之目標之一。 .目前技術： _ R· Agarwal 等學者於”Adaptive Segmentation ofThe digital physiological signals recorded by I can be analyzed and processed by electric contact, which can quickly and efficiently analyze the characteristics and meanings represented by physiological signals, and provide an important reference for medical personnel in diagnosis and interpretation. However, for the Teng signal or other physiological signals, it is still necessary to develop a faster and robust digital signal processing, identification algorithm and classification system designed for long-term monitoring of physiological signals for more accurate and efficient analysis results. And save a lot of human resources and time, and these research work is also one of the goals of the research units at home and abroad. Current technology: _ R· Agarwal and other scholars in "Adaptive Segmentation of
Electroencephalographic Data Using A Nonlinear Energy Operator·” Circuits and Systems, 1999· ISCAS f99. Proceedings of the 1999 IE££ International Symposium on· 4:199-202，1999·之文獻中所提出之方法，係應用非線 性能董法於長時間18電訊號之分段。根據播電訊號之動態 變化，以非線性運算子求得訊號中頻率舆能量之變化狀 態，再利用移動視窗方式對運算子進行可適性閥值設定， 求得頻率舆能量變化之邊界，並以此邊界為雎電訊號各區 參段之分界點。 由於長時間之多項生理訊號處理，必須耗费許多電滕 運算時間，因此利用非線性能量分段法，將生理訊號根據 其頻率舆能董變化，分為數秒至數十秒之時間區段，不但 可以減少電滕運算時間，且各個區段均一性較高，可增加 後續處理之計算精確度。 R· Agarwal等學者於另一篇論文"Automatic EEG analysis during long-term monitoring in ICU·，， 1288875 aElectroencephalographic Data Using A Nonlinear Energy Operator·” Circuits and Systems, 1999· ISCAS f99. Proceedings of the 1999 IE££ International Symposium on 4:199-202, 1999. The method proposed in the literature applies nonlinear energy. Dong Fa is a segment of the long-term 18-signal. According to the dynamic change of the broadcast signal, the nonlinear operator is used to obtain the change state of the frequency 舆 energy in the signal, and then the adaptive window is used to set the adaptability threshold of the operator. The boundary between the frequency and the energy change is obtained, and the boundary is the boundary point of each segment of the electrical signal. Since a lot of physiological signal processing for a long time requires a lot of electric operation time, the nonlinear energy segmentation method is utilized. According to the frequency of the physiological signal, the physiological signal is divided into time segments of several seconds to several tens of seconds, which not only can reduce the operation time of the electric operation, but also has high uniformity of each segment, which can increase the calculation accuracy of subsequent processing. R·Agarwal and other scholars in another paper "Automatic EEG analysis during long-term monitoring in ICU·,, 1 288875 a
II
Electroencephalography and clinical Neurophysiology 107:44-58·此文獻針對長期滕電訊號進行訊號分段、特徵值 擷取，將長時間的鴿電訊號依照其特徵分類，將具有相同 型態之滕電訊號區段分類成群组，以供醫療人員觀察各個 時間區段訊號所屬之群組，便可進一步得知個案之長期滕 電訊號變化狀態。此系統架構應用於加護病房（Intensive Care Unit)，利用上述方法處理病患所董測之大董雎電訊 號，簡化繁複且冗長之檢閱訊號過程，縮短診斷時間舆減 少人力資源，提供醫療人員作為一重要生理狀態參考。 ►然而，上述方法皆針對單一臞電訊號進行分類與處 理，無法針對多軌道之生理訊號進行分析舆分類。且分類 前，未將特徵空間正規化，導致數值較大之特徵項目，在 分類過程中計算相似度之歐基里德距離時，相對地將發揮 較大影響作用，如滕電訊號慢波特徵值；反之，數值較小 之特徵項目，相對地在分類過程中其影響較小，甚至無法 發揮鑑別效能，造成各個特徵項目之不平均現象，結果亦 有可能發生分類錯誤現象。再者，特徵項目若未經過篩選， 選出能夠辨別不同生理訊號型態之特徵項目，亦有可能造 B 成雜訊特徵項目在分類過程中造成干擾，導致分類不精確 現象。 有鑑於此，若能在分類前，先選擇能夠有效地鑑別出 各種型態生理訊號之特徵項目，對分段後的各個生理訊號 區段擷取其特徵值，形成特徵矩陣後，再將特徵空間正規 化’將可大大提升分類之正確性舆精確度。再者，若能將 所有資料處理過程以圖控式介面實現，提供醫療專家操作 使用，並得到完整清楚之各項生理狀態資訊，以及訊號檢 1288875 閱動態視窗，將可大量降低傳統診斷生理訊號之時間舆人 力資源。 【發明内容】 本發明之一目的係在於提供一種長期多項生理訊號 自動處理系統及方法，藉由針對兩軌道膘電訊號、一軌道 之肌電訊號與一軌道之眼電訊號，進行生理訊號的自動前 置處理，包括前置濾波、去雜訊、移除干擾（如董測訊號時 靜電所造成基線（baseline)干擾、病患身艘移動所造成的訊 號變異…等等）。 本發明之再一目的係在於提供一種長期多項生理訊 號自動處理系統及方法，藉由多項生理訊號之自動分段 (segmentation)，係利用非線性能量法將長時間生理訊號分 割為短時間區段，相同區段内的訊號有較為相似的特徵， 便於降低分析上的複雜度舆困難度。 本發明之又一目的係在於提供一種長期多項生理訊 號自動處理系統及方法，藉由對生理訊號區段進行特徵擷 取（feature extraction)。先選定有效之特徵項目後，利用小 波轉換演算法擷取區段内之特徵值，包括雎電訊號之頻率 舆能董、肌電訊號能董、眼球運動、牖電訊號之alpha slow-wave index、theta Sl〇w-wave index等特徵值，形成 特徵矩陣後將其正規化，計算並分析區段内訊號的特徵 值’並將所有區段特徵值儲存於資料庫，可供醫療專家快 速取得個案各個時間下之生理特徵。 本發明之更一目的係在於提供一種長期多項生理訊 號自動處理系統及方法’藉由對所有生理訊號區段進行分 類（classification)工作，根據前一步驟擷取之特徵，利用 1288875 自组式（self-organization)分類法，將具有相似特徵與型態 (pattern)之生理訊號分類成群，並且提供分類後各個群組 之代表樣本訊號舆生理訊號分類時間序列國表，將長達 小時之多項生理訊號以精簡之壓縮型式呈現，藉由此兩項 輸出結果’可判斷出群組内之各項生理特性，並得知病患 在各個時間下所處的狀態舆是否出現異常現象。 本發明之另一目的係在於提供一種長期多項生理訊 號自動處理系统及方法’藉由圈控式操作介面之設計，以 供執行各項步驟之功能，亦可輸入各項處理步驟之重要參 • 數，此外，使用者亦能夠以圈控式訊號顯示介面檢閱經過 各步驟處理後之各項生理訊號，可根據其需求調整各項控 制參數，如振幅尺度、時間轴寬度、特定時段搜尋等功能。 為達上述之目的，本發明係一種長期多項生理訊號自 動處理系統及方法，分析演算法包括四大部分：基於小波 轉換之訊號處理法、非線性能量分段法、區段特徵值擷取 法舆自組式區段分類演算法。而該長期多項生理訊號自動 處理系統係包括：一生理訊號分析系統、一生理訊號分類系 統及一圖控式操控舆顯示介面。該生理訊號分析系統將輸 鲁 入之多項生理訊號進行訊號自動分析；該生理訊號分類系 統將生理訊號分析系統分析後之資料予以分類，並以一麼 縮型式之圖像化方式呈現；而該圖控式操控舆顯示介面係 供執行多項生理訊號分析系統與生理訊號分類系統之訊號 分析、資料分類及圖像化顯示；藉此，利用圈控式操作介 面系統，以長達數小時之受測者大量生理訊號為輸入，經 過上述處理程序後，系統可將具有相似特徵舆型態之生理 訊號時間區段分類成群，並且提供分類後各個群組之代表 1288875Electroencephalography and clinical Neurophysiology 107:44-58·This document performs signal segmentation and eigenvalue extraction for long-term Tengdian signals, and classifies long-term pigeon electric signals according to their characteristics, and will have the same type of Tengdian signal segment. Classified into groups, for medical personnel to observe the group to which each time zone signal belongs, you can further know the status of the long-term telecommunication signal change of the case. The system architecture is applied to the Intensive Care Unit. The above method is used to deal with the Dang Dong Telecom signal of the patient's office, simplifying the complicated and lengthy process of reviewing the signal, shortening the diagnosis time, reducing human resources, and providing medical personnel as a medical staff. An important physiological state reference. ► However, the above methods are classified and processed for a single radio signal, and cannot be analyzed and classified for the physiological signals of multi-track. Before the classification, the feature space is not normalized, resulting in a feature item with a large numerical value. When calculating the Euclid distance of the similarity in the classification process, it will play a relatively large influence, such as the slow wave characteristic of Tengdian signal. On the contrary, the feature items with smaller values have relatively less influence in the classification process, and even can not play the identification efficiency, resulting in the uneven phenomenon of each feature item, and the result may also be classified error phenomenon. Furthermore, if the feature items are not screened, the feature items that can distinguish different physiological signal patterns are selected, and it is also possible that the noise information items cause interference in the classification process, resulting in inaccurate classification. In view of this, if the feature items capable of effectively identifying various types of physiological signals can be selected before classification, the feature values of the segmented physiological signal segments are extracted to form a feature matrix, and then the features are obtained. Spatial normalization will greatly improve the accuracy and accuracy of the classification. Furthermore, if all data processing processes can be implemented in a graphically controlled interface, providing medical experts with operational information, and obtaining complete and clear physiological status information, as well as the signal detection 1288875 reading dynamic window, the traditional diagnostic physiological signals can be greatly reduced. Time 舆 human resources. SUMMARY OF THE INVENTION One object of the present invention is to provide a long-term multi-physical signal automatic processing system and method for performing physiological signals by using two-track 膘 electrical signals, an orbital myoelectric signal, and an orbital ocular electrical signal. Automatic pre-processing, including pre-filtering, de-noising, and removal of interference (such as baseline interference caused by static electricity during the measurement of the signal, signal variation caused by the movement of the patient's body, etc.). A further object of the present invention is to provide a long-term automatic signal processing system and method for multiple physiological signals. By means of automatic segmentation of a plurality of physiological signals, a long-term physiological signal is segmented into short-time segments by a nonlinear energy method. The signals in the same section have similar features, which is convenient to reduce the complexity and difficulty of analysis. Another object of the present invention is to provide a long-term multi-path physiological signal automatic processing system and method for performing feature extraction on a physiological signal segment. After selecting the effective feature item, the wavelet transform algorithm is used to extract the feature values in the segment, including the frequency of the electric signal, the alpha-wave index of the myoelectric signal, the eye movement, and the electric signal. And theta Sl〇w-wave index and other eigenvalues, form the feature matrix and normalize it, calculate and analyze the characteristic value of the signal in the segment' and store all the segment feature values in the database for the medical experts to quickly obtain The physiological characteristics of the case at various times. A further object of the present invention is to provide a long-term multiple physiological signal automatic processing system and method 'by classifying all physiological signal segments, according to the characteristics of the previous step, using 1288875 self-organizing ( Self-organization classification method, which classifies physiological signals with similar characteristics and patterns into groups, and provides representative sample signals of each group after classification, physiological signal classification time series country table, which will be many hours long. The physiological signal is presented in a compact compression format, and the two output results can be used to determine the physiological characteristics of the group and to know whether the patient is in an abnormal state at various times. Another object of the present invention is to provide a long-term multi-path physiological signal automatic processing system and method 'by designing a circle-controlled operation interface for performing various steps, and inputting important steps of each processing step. In addition, the user can also review the physiological signals processed by each step by using the circled signal display interface, and can adjust various control parameters according to the needs thereof, such as amplitude scale, time axis width, and specific time period search. . In order to achieve the above purpose, the present invention is a long-term automatic physiological signal automatic processing system and method, and the analysis algorithm includes four parts: a signal processing method based on wavelet transform, a nonlinear energy segmentation method, and a segment eigenvalue acquisition method. Self-organized segment classification algorithm. The long-term multi-physical signal automatic processing system includes: a physiological signal analysis system, a physiological signal classification system, and a graphical control display interface. The physiological signal analysis system automatically analyzes a plurality of physiological signals input into the body; the physiological signal classification system classifies the data analyzed by the physiological signal analysis system, and presents the image in a reduced format; The graphically controlled display interface is used to perform signal analysis, data classification and graphical display of a plurality of physiological signal analysis systems and physiological signal classification systems; thereby, the use of a circled operation interface system for up to several hours The tester has a large number of physiological signals as inputs. After the above processing procedure, the system can classify the physiological signal time segments having similar characteristic 舆 patterns into groups, and provide representatives of each group after classification 1288875
I i 樣本訊號舆生理訊號分類時間序列圓表，以精簡之壓縮型 式呈現，並將處理後的生理訊號狀態建立成快取之多項生 理訊號資料庫，可供使用者即時査詢受測者特定時間之生 理訊號與分析後之特徵資料者。 【實施方式】 本發明之技術内容舆操作方式，將以圖示配合實際範 例，詳細說明其操作過程優點舆特點。 請參閱第一至十圈所示，本發明係為一種長期多項生 理訊號自動處理系統，該長期多項生理訊號自動處理系統 藝包括有： 一生理訊號分析系統A，將輸入之多項生理訊號4進 行訊號自動分析；該生理訊號分析系統A係包括： 一訊號處理單元6 1，該訊號處理單元6 1係包含前 置處理單元6 1 1、小波轉換處理6 1 2。 一生理訊號分段單元6 4，該生理訊號分段單元6 4 係以非線性能量分段準則6 1 3進行訊號分段。 一特徵擷取單元6 2，該特徵擷取單元6 2係將訊號 處理舆訊號分段處理後之多項生理訊號，擷取區段特徵值 • 621，並將各區段之各項生理特徵值建立區段特徵值矩 陣 6 2 2。 一生理訊號分類系統B，將生理訊號分析系統A分析 後之資料予以分類，並以一壓縮型式之圖像化方式呈現； 該生理訊號分類系統B係包括： 一自組式分類演算單元63，該自組式分類演算單元 6 3係以特徵值矩陣為輸入，進行模糊分類演算法，將具 有相似特徵值之區段分類為同一群組，特徵值相異之區段 1288875 分為不同群组β 一系統輸出7，該系統輸出7之輸出内容包含： 一提供分類後各個群組之代表樣本訊號7 1，而一個 具有相同特徵值之群組，其多項生理訊號以一組代表樣本 訊號表示之。 一以精簡之壓縮型式呈現之生理訊號分類時間序列圖 表7 2,以供使用者快速檢閱受試者之生理狀態變化情況。 一圖控式操控舆顯示介面C，執行多項生理訊號分析 g 系統Α舆生理訊號分類系統Β之訊號分析、資料分類及圈 像化顯示者；該圈控式操控舆顯示介面C係包括： 一圖控式操控介面5 1，係供執行多項生理訊號分析 系統A舆生理訊號分類系統B之各項處理動作。 一圈控式訊號顯示介面5 2，係以可調控式之移動視 窗，供使用者檢閱經過處理後之各項生理訊號，並根據其 需求調整各項控制參數者。 另，該長期多項生理訊號自動處理系統係進一步包含 有一多項生理訊號資料庫8，以供儲存所有處理後之生理 ^ 訊號特徵舆資訊者。 本發明之長期多項生理訊號自動處理方法，其包括下 列步驟： 一、 將多項生理訊號做訊號處理； 二、 擷取其區段特徵值； 三、 建立區段特徵值矩陣； 四、 進行自組式分類演算； 五、 提供分類後各個群组之代表樣本訊號及生理訊衆分 類時間序列圖表；以及 1288875 % 秦 六、將所有處理後之生理訊號特徵與資訊儲存於一多項 生理訊號資料庫。 首先如第一圖，受試者經由感測電極裝置1操取電生 理訊號，再經由一訊號放大模組2與一訊號感測棋组3將 董測到之各項生理訊號4輸入至一長期多項生理訊號自動 分類系統5。透過一圖控式操作介面5 1與一圈控式訊號 顯示介面5 2依使用者需要，調整各項輸入舆輸出參數， 而多項生理訊號資料將透過各項參數之設定經由一長期多 • 項生理訊號自動分類演算法6處理後，提供使用者一系統 輸出7，並將所有處理後之生理訊號特徵舆資訊，儲存於 一多項生理訊號資料庫8，供使用者及時檢閱舆了解受試 者之各項生理訊號狀態。 如第二圈，為長期多項生理訊號自動分類演算法之詳 細處理步驟，以四個通道之多項生理訊號4，包括電極位 置 Fpz-Cz 4 1 與 Ρζ·Οζ 4 2 之滕電訊號 (electroencephalogram 9 EEG)、頦部肌電訊號 4 3 (eleetromyogram，EMG)舆水平眼電訊號 4 4 (horizontal 鲁 electrooculogram，EOG)為輸入訊號，經由訊號處理單元 6 1，包括了移除基線干擾、前置濾波、雜訊偵測以及閥 值設定。 如第三圖，為第二圈中非線性能量分段舆擷取區段特徵值 之示意圈。經過前置處理後之四軌道訊號，將同時被切割 成擬穩態區段，且區段時間必須大於3秒，以確保區段特 徵擷取之意義。利用非線性能量分段法則，依照訊號之振 幅舆頻率變化做分段，此分段標記同時用於四段生理訊號。 12 1288875 第四圈為第二圖舆第三圖中自组式分類演算法之詳 細流程圈，首先，我們先將區段特徵矩陣A 1經過「正規 化A 2 (normalization)」處理，將每個特徵的數值範圍進 行線性調整，使其範圍落於〇舆1之間。正規化的過程能 夠讓各項特徵之數值範圍不會相差太大，若未經正規化處 理，數值較大的特徵項目之影響將會遠大於數於較小之特 徵項目，造成分群時各項特徵值所佔之比重不平均的情況 發生，將造成數值小的特徵項目無法發揮其分類作用。自 组式分類演算法步驟如下說明： 階段一 •消除雜訊區段A3 :由上述所設定之腦電訊號300μν臨 界振幅，移除超過此 振幅的EEG區段，以降低雜訊對於分群結果之影響。 • 分群為心個群聚Α4 :以所有區段特徵（消除I i sample signal 舆 physiological signal classification time series round table, presented in a compact compression format, and the processed physiological signal state is established into a cache of multiple physiological signal databases, allowing the user to instantly query the subject for a specific time The physiological signal and the characteristic data after analysis. [Embodiment] The technical contents and operation modes of the present invention will be described in detail with reference to actual examples, and the advantages and characteristics of the operation process will be described in detail. Referring to the first to tenth laps, the present invention is a long-term multi-physical signal automatic processing system. The long-term physiological signal automatic processing system includes: a physiological signal analysis system A, which inputs a plurality of physiological signals 4 The signal analysis system A includes: a signal processing unit 161, the signal processing unit 6.1 includes a pre-processing unit 611, and a wavelet conversion process 612. A physiological signal segmentation unit 64, the physiological signal segmentation unit 64 performs signal segmentation with a nonlinear energy segmentation criterion 613. A feature capturing unit 6 2 is configured to process the plurality of physiological signals after the signal processing of the signal, and extract the segment feature value • 621, and the physiological characteristic values of each segment A segment feature value matrix 6 2 2 is established. A physiological signal classification system B classifies the data analyzed by the physiological signal analysis system A and presents it in a compressed version of the image; the physiological signal classification system B includes: a self-organized classification calculation unit 63, The self-organized classification calculation unit 6 3 uses the eigenvalue matrix as input to perform the fuzzy classification algorithm, and classifies the segments with similar eigenvalues into the same group, and the segment 1288875 with different eigenvalues is divided into different groups. Β-system output 7, the output of the system output 7 includes: a representative sample signal 7 of each group after classification, and a group having the same characteristic value, the plurality of physiological signals represented by a set of representative sample signals It. A physiological signal classification time series chart presented in a reduced compression format is shown in Table 7 2 for the user to quickly review changes in the physiological state of the subject. A graphic control device 舆 display interface C, performing a plurality of physiological signal analysis g system Α舆 physiological signal classification system 讯 signal analysis, data classification and circle image display; the circle control operation 舆 display interface C system includes: The graphical control interface 5 1 is for performing various processing actions of the physiological signal analysis system A and the physiological signal classification system B. A circle of control signal display interface 52 is a controllable mobile window for the user to review the processed physiological signals and adjust the control parameters according to their needs. In addition, the long-term physiological signal automatic processing system further comprises a plurality of physiological signal databases 8 for storing all processed physiological signals. The long-term multi-path physiological signal automatic processing method of the invention comprises the following steps: 1. processing a plurality of physiological signals as signals; 2. drawing a segment feature value; 3. establishing a segment feature value matrix; 4. performing self-grouping Classification calculus; 5. Provide representative sample signal and physiological population classification time series chart of each group after classification; and 1288875 % Qin Six, store all processed physiological signal features and information in a plurality of physiological signal database . First, as shown in the first figure, the subject operates the electrophysiological signal through the sensing electrode device 1, and then inputs the physiological signals 4 detected by Dong through a signal amplification module 2 and a signal sensing chess group 3. Long-term multiple physiological signal automatic classification system 5. Through a picture-controlled operation interface 5 1 and a circle-controlled signal display interface 5 2, according to the needs of the user, various input and output parameters are adjusted, and a plurality of physiological signal data are set through a long-term and multiple items through various parameters. After the physiological signal automatic classification algorithm 6 is processed, the user provides a system output 7 and stores all processed physiological signal characteristics and information in a plurality of physiological signal database 8 for the user to review in time to understand the test. The physiological signal status of the person. For example, the second lap is a detailed processing procedure for the long-term automatic classification algorithm of multiple physiological signals, with a plurality of physiological signals 4 of four channels, including electrode positions Fpz-Cz 4 1 and en·Οζ 4 2 electrical signal (electroencephalogram 9 EEG), eleetromyogram (EMG) 舆 horizontal EO signal 4 4 (horizontal ruth electrooculogram, EOG) is an input signal, through the signal processing unit 161, including removal of baseline interference, pre-filtering , noise detection and threshold setting. As shown in the third figure, a schematic circle of segment feature values is extracted for the nonlinear energy segment in the second circle. The pre-processed four-track signal will be cut into a quasi-steady-state segment at the same time, and the segment time must be greater than 3 seconds to ensure the significance of the segment feature. Using the nonlinear energy segmentation rule, segmentation is performed according to the amplitude of the signal, and the segmentation mark is used for the four-segment physiological signal. 12 1288875 The fourth circle is the detailed flow circle of the self-organized classification algorithm in the second picture 舆 third picture. First, we first process the segment feature matrix A 1 by “normalization A 2 (normalization)”. The range of values of the features is linearly adjusted so that the range falls between 〇舆1. The process of normalization can make the numerical range of each feature not much different. If it is not normalized, the influence of the feature items with larger values will be much larger than that of the smaller feature items, resulting in various groups. When the proportion of the eigenvalues is not uniform, the feature items with small values cannot be used for classification. The steps of the self-organizing classification algorithm are as follows: Stage 1 • Elimination of the noise segment A3: The EEG segment exceeding the amplitude is removed by the critical amplitude of the EEG signal 300 μν set above to reduce the noise for the clustering result. influences. • Grouping into groups of hearts 4: with all segment features (eliminate
Artifact區段特徵後）為輸入，利用FCM法分群，並設定 「初始群聚個數」幻 =10,使數目大於「最终群聚個 數」Λ/ =6。其原因是由於在這些特徵内，仍然有許多 因雜訊干擾之區段特徵，在此稱之為特異（outlier)特 徵，這些特徵將嚴重影響最終分群結果，因此將幻值設 定高於心值之目的，是希望能夠將特異特徵分離出來， 並消除這些以奇異特徵向量為主之群聚，而值即為睡 眠等級之個數（睡眠等級1〜4、REM睡眠期與清醒期）。 階段二 •消除特異區段並建立特異區段索引A5:完成階段一 之第二個步驟後，可得到10個群聚。在此1❶個群聚中， 若有群聚其包含之區段數少於，即令其為特異群 13 1288875 聚，此群聚所包含之區段變令其為特異區段。設定 為所有區段總數之1°/❶，這些特異區段之特徵值將從特 徵矩陣中移除，且不參舆後續處理流程。在臨床診斷 上，這些特異區段内可能存在著具有特殊意義的暫態特 徵訊號，但在睡眠等級判別方面，最主要的依據是由長 時間背景波段的變化（如alpha、beta等節律變化），來判 定其睡眠等級，除去低於總數1 %的特異區段，除了可 以降低這些區段分群上的影響，且在個數方面也是在可 接受範圍之内。 齡 •將群聚分為兩大組A6:此步驟是將剩餘之區段分為兩 大群組’其主要目的 是將分布狀態遠離大多數群聚之群聚，分為獨立之群 組’避免這些疏遠的群聚在下一步称（merge clusters) 造成影響。利用計算每個群聚之群聚相互距離 (inter-cluster distance)，將大於平均距離2倍標準差之 群聚分為一個群組；其餘的群聚分為另一個群組，群聚 相互距離定義為某一群聚中心與其餘各群聚中心之距 離之平均值。 >·融合各組内群聚A7:經過上一步驟，將所有群聚分為 兩大群取，此步驟將 個別對兩群组内之群聚作合併之工作，主要目的是要將 過於相近的群聚合併，以減少過多之群聚數目，使其較 為接近最终群聚數目。群聚合併演算法如下： then merge the kx and ky clusters kxKjky^kx 1288875After the Artifact segment feature is input, the FCM method is used to group and set the number of "initial clusters" to be imaginary = 10, so that the number is greater than the "final cluster number" Λ / = 6. The reason for this is that, among these features, there are still many segment features due to noise interference, referred to herein as outlier features, which will seriously affect the final clustering result, so the magic value is set higher than the heart value. The purpose is to separate the specific features and eliminate these clusters with singular feature vectors, and the value is the number of sleep levels (sleep level 1~4, REM sleep period and awake period). Phase 2 • Eliminate specific segments and establish a specific segment index A5: After completing the second step of phase one, 10 clusters are obtained. In this group of clusters, if there are clusters containing less than the number of segments, it is made to be a specific group 13 1288875, and the segment included in the cluster becomes a specific segment. Set to 1°/❶ of the total number of all segments, the feature values of these specific segments will be removed from the feature matrix and will not be referenced in the subsequent processing. In clinical diagnosis, there may be transient characteristics signals with special significance in these specific segments, but in terms of sleep level discrimination, the most important basis is the change of long-term background band (such as alpha, beta and other rhythmic changes). To determine the sleep level, remove the specific segments below 1% of the total, in addition to reducing the impact on the segmentation of these segments, and in terms of the number is also within the acceptable range. Age • Divide the group into two groups A6: This step is to divide the remaining segments into two groups. The main purpose is to separate the distribution from most clusters and divide them into separate groups. Avoid these alienated clusters in the next step (merge clusters). By calculating the inter-cluster distance of each cluster, the clusters larger than the standard deviation of 2 times the standard deviation are grouped into one group; the remaining clusters are divided into another group, and the clusters are separated from each other. Defined as the average of the distance between a cluster center and the remaining cluster centers. >·Integration of cluster A7 in each group: After the previous step, all clusters are divided into two groups. This step will combine the work of the groups in the two groups. The main purpose is to be too close. The clusters are aggregated to reduce the number of clusters that are too close to the final population. The group aggregation and algorithm are as follows: then merge the kx and ky clusters kxKjky^kx 1288875
Repeat until the above criterion is not satisfied. 其中以)為任兩群聚中心t與心之歐基里德距離，^ 為所有兩兩成對群聚中心距離之平均β 為可調變之 參數，經由實驗此分析流程所得到的結果，發現將此參 數設定為== 40%將得到最佳的群聚合併結果。此一 參數的設定並無最佳化方法，因為群聚合併本身就是一 項困難的問題，只能靠經驗法則以及觀察群组合併後結 果來作參數調整。 •剩餘群聚 > 妗A 8 :經過上述步驟後，可得到經過合 併後的群聚數目t，若t小於h，則以目前的群聚中心 為初始輸入值執行FCM演算法，並以L為分群數目；若 幻大於心，則將目前的群聚中心以其資料量為依據，依 序取出前個群聚中心，作為初始輸入值執行fcm演算 法，並以心為分群數目。此步驟主要目的為讓資料群的 群聚數目不超過最終分群個數b(即減少群聚數至心A 9) ’若實際分群數目小於心，表示此個案再睡眠時所 處的睡眠等級較少，例如：呼吸窒息患者也許在整個晚 上的睡眠等級只有包含清醒與等級1、2的淺睡期，所以 此項假設合理。 • FCM分群（利用目前群聚中心為輸入）A 1 〇 :在上一步 称有提到’以合併後的群聚中心為初始值作為檢入，此 目的是因為經過處理舆合併後的群聚中心將具有一定 的代表性，且其分群結果亦會越接近整個資料中實際群 聚的分布狀態，再作一次FCM分群之目的是要將合併後 的群聚其分布邊緣的資料（遠離群聚中心的資料），在重 新依照舆群聚中心間的距離，指定其所屬的確切群聚。 15 1288875 階段三 此步驟的概念和擅段二相似，若所有資料中已經沒有特 異區段，則結束演算法，若還存在特異區段，則必須再 次消除特異區段A 1 1，再判斷是否有任何雜訊干擾區 段A12,並重新利用FCM分群A 1 3(以目前的群聚 中心為輸入）。 如第五圈到第七圈，多項生理訊號自動分類ACSP之 系統輸出7包含下列三項：第一是「展示出對於這5個類 別分別具有其代表性的區段生理訊號」（即提供分類後各個 . 群組之代表樣本訊號7 1 )，也就是選擇舆群聚中心特徵值 相近之區段，如第五圖；第二是「長期生理訊號區段之分 類時間序列圖表」（即以精簡之壓縮型式呈現之生理訊號分 類時間序列圈表7 2 )，每個區段將依照其所屬群聚，標上 特定群聚所代表之顏色，而區段在時間序列上的長度將和 其實際長度成比例，如第六圈；第三是「長期epoch生理 訊號區段之分類時間序列」，同第二項，由每個epoch(30 秒之區段）内之區段群聚，決定此epoch之所屬群聚，每個 epoch將依照其所屬群聚，標上特定群聚所代表之顏色， B 而區段在時間序列上的長度將和其實際長度成比例，如第 七圖。以上描述之三項輸出，將可完整呈現並描述長時間 生理訊號之變化。利用Physionet睡眠生理資料庫之個案 (檔案編號：st7022j0.rec)整晚之睡眠生理訊號為輸入，所 得到之結果如下： 輸出1 :第五圈為經過自組式分類演算法，將個案所 有生理訊號分成5個類別後，取出舆群聚中心具有最相似 特徵之生理訊號區段，為各個群聚Type 1-Type 5之代表 16 1288875 性區段，由圈可知’不同群聚Type 1 - Type 5之代表訊號 區段，其訊號特徵皆不同，如Type 5舆其他型式之最大不 同處，為其高EMG振幅以及高BeU頻帶能量，且有較多 之眼球運動，判斷個案狀態為為清醒期。 輸出2:第八圈為一時間為6分鐘之區段生理訊號變 化之範例，不同顏色之區域代表不同群聚（型態）之區段， 舉例來說，圈中時間序列輛上所出現「綠色」之區域，代 表這些區段的生理訊號為「Type 3」之型式，而其後所有 出現之綠色區域將表不此區段為同一型態之生理訊號區 春 段，換句話說，這些綠色區段將具有相同之特徵舆型態， 同理，其他顏色也代表為另一種型態之生理訊號區段，如 第八圈之紅色舆紫色區段。長達近8個小時之分類時間序 列如第六圈所示，此個案之資料經處理後共包含了 4,786 個生理訊號區段，其中有43個區段經由系統判斷為特異/ 雜訊干擾之特異區段。 輸出3 :利用epoch為群聚單位之分類時間序列如第 七圖所示，由此時間序列可看出個案在長期睡眠過程中， 其生理訊號變化以及不同生理狀態所佔之比例多募。 _ 如第九圖，為本發明之參數化圈控式操控介面51。 上述所執行之長期多項生理訊號自動分類演算法6之處理 步驟，皆可由此介面執行，包括前置濾波、非線性能董分 段、特徵值擷取舆自组式分類。各項步驟執行後，系統將 會以彈跳視窗方式，顯示出如第十圖之圈控式訊號顯示介 面5 2，供使用者檢閱經過處理後的訊號。使用者可調整 介面之各項調控參數9，包括了振幅尺度參數9 1、時間 定位參數9 2以及滑鼠指標位置9 3。 17 1288875 【圈式簡單說明】 第一圖係本發明長期多項生理訊號自動分類系統之整髏實 施舆配置式意圖。 第二圖係本發明系統之主要演算步驟流程圈。 第三圖係本發明利用非線性能量法進行訊號分段舆擷取特 徵值之示意圖。 第四圖係本發明基於模糊C-Means之自組式分類演算法主 要步麻流程圈。 第五圈係本發明之系統輸出一：各群聚之代表樣本訊號。 第六圖係本發明之系統輸出二：區段分類時間序列圖表。 第七圖係本發明之系統輸出三·· epoch分類時間序列圖表。 第八圈係本發明之6分鐘區段分類時間序列圈表示意圈。 第九圈係本發明之參數化圈控式操控介面圈。 第十圈係本發明之圖控式訊號顯示介面圈。 【主要元件符號說明】 1、 感測電極裝置 2、 訊號放大模组 3、 訊號感測模組 4、 各項生理訊號 4 1、Fpz-Cz之牖電訊號 4 2、Pz-Oz之J8S電訊號 43、水平眼電訊號 φ 4 4、頦部肌電訊號 A、生理訊號分析系統 5、 長期多項生理訊號自動分類系統 5 1、圈控式操作介面 5 2、圈控式訊號顯示介面 6、 長期多項生理訊號自動分類演算法 61、訊號處理單元 611、前置處理單元 6 1 2、小波轉換處理 6 1 3、非線性能量分段準則 6 2、特徵擷取單元 1288875 6 2 1、擷取區段特徵值 6 2 2、建立區段特徵值矩陣 6 4、生理訊號分段單元 B、生理訊號分類系統 6 3、自组式分類演算單元 7、 系統輸出 7 1、代表樣本訊號 7 2、分類時間序列圈表 8、 多項生理訊號資料庫 9、 調控參數 9 1、振幅尺度參數 9 2、時間定位參數 • 93、滑鼠指標位置 A 1、區段特徵矩陣 A 2、正規化 A3、消除雜訊區段 A 4、FCM分群為ki個群聚 A5、消除特異區段並建立特異區段索引 A 6、將群聚分為兩大組 A7、融合各組内群聚 A8、剩餘群聚>kf? A9、減少群聚數至kf A 1 〇、FCM分群（利用目前群聚中心為輸入） φ All、再次消除特異區段 A1 2、是否有任何雜訊干擾區段 A 1 3、FCM 分群 19Repeated to the above criterion is not satisfied. Among them is the distance between the two groups of polycenter t and the Euclid distance of the heart, ^ is the parameter that the average β of all pairs of pairs of cluster centers is adjustable, via Experiment with the results of this analysis process and found that setting this parameter to == 40% will result in the best group aggregation and results. There is no optimization method for this parameter setting, because group aggregation is a difficult problem in itself. It can only be adjusted by the rule of thumb and the result of observation group combination. • Remaining clusters > 妗A 8 : After the above steps, the number of merged clusters t can be obtained. If t is less than h, the FCM algorithm is executed with the current cluster center as the initial input value, and L For the number of clusters; if the magic is greater than the heart, the current cluster center is based on the amount of data, and the previous cluster center is taken out in order, and the fcm algorithm is executed as the initial input value, and the number of clusters is taken as the heart. The main purpose of this step is to make the number of clusters of the data group not exceed the number of final clusters b (ie, reduce the number of clusters to the heart A 9). 'If the actual number of clusters is less than the heart, it means that the sleep level of this case is more sleep. Less, for example, patients with respiratory asphyxia may have a sleep level of only awake and grade 1, 2 throughout the night, so this assumption is reasonable. • FCM grouping (using the current cluster center as input) A 1 〇: In the previous step, it was mentioned that 'the combined value of the clustered center is taken as the initial value. This is because the processed and merged clusters are processed. The center will have a certain representativeness, and the clustering result will be closer to the actual clustering distribution in the whole data. The purpose of making another FCM clustering is to gather the combined information of the distributed edges (away from clustering). The information of the center), in accordance with the distance between the center of the 舆 cluster, specify the exact cluster to which it belongs. 15 1288875 Stage 3 The concept of this step is similar to that of the second paragraph. If there is no specific section in all the data, the algorithm ends. If there is a specific section, the specific section A 1 1 must be eliminated again. Any noise interferes with segment A12 and reuses FCM cluster A 1 3 (entered with the current cluster center). For the fifth to seventh laps, the system output 7 of multiple physiological signal automatic classification ACSPs contains the following three items: The first is to "show the physiological signals of the segments that are representative of the five categories" (ie, provide classification The representative sample signal of each group is 7 1 ), that is, the section where the eigenvalues of the 舆 cluster center are similar, such as the fifth map; the second is the classification time series chart of the long-term physiological signal section (ie, Streamlined compression pattern presented by the physiological signal classification time series loop table 7 2), each segment will be grouped according to its group, marked with the color represented by the specific cluster, and the length of the segment in the time series will be The actual length is proportional to the sixth circle; the third is the "longitudinal time series of long-term epoch physiological signal segments", and the second item is determined by the clustering of segments within each epoch (30-second segment). The clusters to which they belong, each epoch will be grouped according to its group, marked with the color represented by the particular cluster, B and the length of the segment in time series will be proportional to its actual length, as shown in the seventh figure. The three outputs described above will fully present and describe changes in long-term physiological signals. The Physionet sleep physiology database case (file number: st7022j0.rec) is used as input for the whole night's sleep physiological signal. The results obtained are as follows: Output 1: The fifth lap is a self-organized classification algorithm that will treat all the physiology of the case. After the signals are divided into five categories, the physiological signal segments with the most similar characteristics in the cluster center are taken out, which are 16 1288875 segments of each group Type 1-Type 5, and the circle is known as 'different cluster Type 1 - Type The representative signal segment of 5 has different signal characteristics, such as the biggest difference between Type 5 and other types, its high EMG amplitude and high BeU band energy, and there are more eye movements, and the case status is determined as the awake period. . Output 2: The eighth circle is an example of a physiological signal change in a segment with a time of 6 minutes. The regions of different colors represent segments of different clusters (types). For example, the time series on the circle appears. The green area represents the physiological signal of these segments as "Type 3", and all subsequent green areas will indicate that the segment is the same type of physiological signal zone, in other words, these The green segments will have the same characteristic 舆 type, and the other colors also represent the physiological signal segments of the other type, such as the red 舆 purple segment of the eighth circle. The classification time series of up to 8 hours is shown in the sixth circle. The data of this case has been processed to include 4,786 physiological signal segments, of which 43 segments are judged as specific/noise interference by the system. Specific segment. Output 3: The chronological sequence using epoch as the clustering unit is shown in Figure 7. From this time series, it can be seen that during the long-term sleep, the physiological signal changes and the proportion of different physiological states are increased. _ As shown in the ninth figure, the parameterized circle-controlled manipulation interface 51 of the present invention. The processing steps of the long-term multi-path physiological signal automatic classification algorithm 6 performed above can be performed by the interface, including pre-filtering, non-linear energy segmentation, eigenvalue acquisition, and self-group classification. After the steps are executed, the system will display the circled signal display interface 52 of the tenth figure in the pop-up window for the user to review the processed signal. The user can adjust various control parameters 9 of the interface, including the amplitude scale parameter 9 1 , the time positioning parameter 9 2 , and the mouse pointer position 9 3 . 17 1288875 [Simple description of the circle] The first picture is the intention of the implementation of the long-term multiple physiological signal automatic classification system of the present invention. The second figure is the main calculation step flow circle of the system of the present invention. The third figure is a schematic diagram of the present invention using the nonlinear energy method for signal segmentation extraction characteristic values. The fourth figure is based on the fuzzy C-Means self-organizing classification algorithm. The fifth circle is the system output of the present invention: a representative sample signal of each group. The sixth figure is the system output 2 of the present invention: a segment classification time series chart. The seventh figure is a system output of the present invention. The epoch classification time series chart. The eighth circle is the 6 minute segment classification time series circle of the present invention representing the circle of interest. The ninth circle is a parametric circle-controlled manipulation interface ring of the present invention. The tenth circle is the graphic control signal display interface circle of the present invention. [Main component symbol description] 1. Sensing electrode device 2, signal amplifying module 3, signal sensing module 4, various physiological signals 4 1 , Fpz-Cz, telecommunication signal 4 2, Pz-Oz J8S telecommunication No. 43, horizontal EO signal φ 4 4, sacral muscle signal A, physiological signal analysis system 5, long-term multiple physiological signal automatic classification system 5 1, circle-controlled operation interface 5 2, circle-controlled signal display interface 6, Long-term multiple physiological signal automatic classification algorithm 61, signal processing unit 611, pre-processing unit 6 1 2, wavelet conversion processing 6 1 3, nonlinear energy segmentation criterion 6 2, feature extraction unit 1288875 6 2 1, capture Segment feature value 6 2 2. Establish segment feature value matrix 6 4. Physiological signal segmentation unit B, physiological signal classification system 6 3. Self-organized classification calculation unit 7, system output 7 1, representative sample signal 7 2. Classification time series circle table 8, multiple physiological signal database 9, regulation parameters 9 1 , amplitude scale parameter 9 2, time positioning parameters • 93, mouse index position A 1 , segment feature matrix A 2, normalized A3, elimination Noise section A 4 The FCM group is ki cluster A5, the specific segment is eliminated and the specific segment index A 6 is established, the cluster is divided into two groups A7, the cluster A8 in the fusion group, and the remaining clusters > kf? A9, Reduce the number of clusters to kf A 1 〇, FCM clustering (using the current cluster center as input) φ All, eliminate the specific segment A1 again 2. Whether there is any noise interference segment A 1 3, FCM group 19