失智症對於健康或經濟都有很大的影響，診斷繁雜不易、退化速度快，且目前缺乏有效藥物治療失智症，根據世界衛生組織WHO的估計全球失智症人口2019年已超過五千萬人，失智症相關成本每年約一兆美元。根據台灣失智症協會指出，在台灣65歲以上的老人每12人即有1位為失智症患者，而80歲以上每5人即有1位為失智症患者。
目前失智症無法透過單一檢測進行診斷，需透過問診、認知功能檢測、神經學檢查、抽血檢查，以及電腦斷層（CT）與核磁共振（MRI）等多項檢測綜合評估才能進行診斷，失智症檢測流程不僅耗時、成本高且包含侵入式檢測，更大大的局限檢測地點，必須經歷漫長痛苦歷程往返醫院。過去前人的研究指出失智症與血液循環及供血效率有關連性，且在腦部相關的中風疾病與脈波參數有相關影響。因此本實驗室期望能利用手腕橈動脈(BPW)及手指末梢(PPG)的脈波量測得出其供血相關頻域參數，並結合機器學習對失智症做出即時、客觀且非侵入的早期偵測及社區場域失智症風險(心智障礙程度)評估。
本研究主要架構是穿戴式裝置結合脈波參數分析，並結合機器學習分類，使用本實驗室自行開發的穿戴式脈波量測系統進行失智症及社區場域數據的深入探討分析，透過脈波量測系統包含的設備有動脈血壓波形模組(BPW)及光血容積訊號儀(PPG)，分別量測動脈端與末梢端，將BPW與PPG量測到的時域脈波資料轉成頻域參數進行後續計算，並進行八種機器學習模型找出最適合的分類器。
為了落實失智症早期偵測及提供客觀評估指標，本實驗基於目前仁愛醫院失智症門診現況，在不影響患者就醫權益的前提下對失智症持續擴大深入收案，並嘗試往社區場域進行量測分析，深入了解不同類型的社區場域之失智症風險及其心智障礙程度影響，社區場域包括清幽類型社區場域及文教類型社區場域，使穿戴式裝置應用於更廣大層面的使用者，同時對失智症及心智障礙程度相關主題有所幫助。
實驗結果表示脈波頻域參數在1.失智症組、不同社區場域(清幽類型及文教類型)及對照組之間在參數變異度的統計上有顯著差異2.不同病程(輕、中、重度)的失智症組之間略有差異3.失智症程度與心智障礙程度之間相似的表現趨勢，進一步應用於社區心智健康程度分析。透過建立機器學習模型，找出具有最好分類能力的演算法，最後再進行每筆數據實際測試，其中高斯貝葉斯方法(Gaussian Naive Bayes, GNB)演算法與K-近鄰演算法(K-Nearest Neighbor Classification,KNN)演算法分別對重度失智症BPW及輕度失智症PPG偵測具有一定程度的分類能力，ROC曲線下面(AUC)分別為0.66及0.6，而失智症與健康組間的差異也在觀察完統計結果後進行機器學習訓練，線性區別分析(Linear Discriminant Analysis, LDA)演算法於BPW或PPG也都具有一定程度的分類能力，ROC曲線下面(AUC)約為0.7，並利用此模型將兩社區場域之每筆數據進行實際測試，透過MMSE量表劃分心智障礙程度，並將社區場域每筆數據之模型測試結果比對分析，於整體社區場域心智健康狀態評估也有良好的結果，實際應用層面，BPW(「無」心智障礙判斷準確率為77.78%)比PPG更適合應用於早期偵測，而中期心智障礙檢測應用PPG(「中度」心智障礙判斷準確率為85.71%)比BPW更適合，代表此穿戴式系統結合機器學習模型能很好應用在社區量測數據上，且穿戴式裝置結合以上應用於機器學習的判別上也具有一定程度的解析能力。
本實驗室開發之裝置經過改良與結合機器學習，不僅可以實際運用在臨床分析，也能在社區居家環境中使用，且具有一定程度的解析能力，在臨床應用上能夠及時提供更具客觀性的評估，同時讓病症有更仔細的分級。協助醫師在社區居家環境達到早期偵測與早期預防的目標。

Dementia has a great impact on health and the economy. Diagnosis is complicated and difficult, degeneration is fast, and there is currently a lack of effective drugs to treat dementia. According to the World Health Organization WHO estimates that the global population of dementia has exceeded 5,000 in 2019 For 10,000 people, the cost of dementia is about one trillion U.S. dollars per year. According to the Taiwan Dementia Association, in Taiwan, 1 out of 12 people over 65 years old is a patient with dementia, and 1 out of every 5 people over 80 years old is a patient with dementia.
Dementia can't be diagnosed by a single detection. It needs to be diagnosed through
interrogation, cognitive function testing, neurological examination, blood test, and comprehensive evaluation of multiple tests such as computer tomography (CT) and magnetic resonance (MRI) etc. Dementia detection process is not only time-consuming, and includes invasive detection. Previous studies have pointed out that dementia is related to blood circulation and blood supply efficiency, and stroke diseases related to the brain have related effects on pulse wave parameters. Therefore, our laboratory hopes to use the pulse wave measurements of the radial wrist (BPW) and the tip of the finger (PPG) to obtain the blood supply-related frequency domain parameters, and combine machine learning to make an immediate, objective and non-invasive diagnosis of dementia. Assessment of the risk of dementia in the community field (degree of mental functions) early.
The main framework of this research is wearable device combined with pulse wave parameter analysis, combined with machine learning classification, and the wearable pulse wave measurement system developed by our laboratory is used to conduct in-depth discussion and analysis of dementia and community field data. The wave measurement system includes the arterial blood pressure waveform module (BPW) and the photoblood volume signal (PPG), which measure the arterial end and the peripheral end respectively, and convert the time-domain pulse wave data measured by BPW and PPG into The frequency domain parameters are subsequently calculated, and eight machine learning models are performed to find the most suitable classifier.
In order to implement the early detection of dementia and provide objective evaluation indicators, this experiment is based on the current situation of the dementia clinic of Ren'ai Hospital, and continues to expand the admission of dementia cases without affecting the rights and interests of patients, and try to go to the community. To conduct measurement and analysis in different types of community fields to understand the risk of dementia and the impact of mental disability in different types of community fields. The community fields include quiet community fields and cultural and educational community fields, so that wearable devices can be used in a wider range of applications. At the same time, it is helpful for the subjects related to the degree of dementia and mental disability.
The experimental results show that the pulse wave frequency domain parameters are 1. There are statistically significant differences in the parameter variability between the dementia group, different community fields and the control group. 2. There are slight differences between the dementia groups of different courses of disease 3. The similar manifestation trend between the degree of dementia and the degree of mental disability is further applied to the analysis of the mental health of the community. Through the establishment of a machine learning model, find the algorithm with the best classification ability, and finally perform the actual test of each data. Among them, the Gaussian Naive Bayes algorithm (GNB) and the K-Nearest Neighbor Classification algorithm (KNN) are respectively The detection of BPW and mild dementia PPG has a certain degree of classification ability. The ROC curve (AUC) is 0.66 and 0.6 respectively. The difference between dementia and healthy groups is also machined after observing the statistical results. For learning and training, the Linear Discriminant Analysis algorithm (LDA) has a certain degree of classification ability in BPW or PPG, and the ROC curve (AUC) is about 0.7. Using this model, each data in the two community fields is actually tested , Through the MMSE scale to classify the degree of mental disability, and compare and analyze the model test results of each data in the community field. The evaluation of the mental health status of the overall community field also has good results. At the practical application level, BPW ("No" Mind Disability judgment accuracy rate is 77.78%) is more suitable for early detection than PPG, and mid-term mental disability detection application PPG ("moderate" mental disability judgment accuracy rate is 85.71%) is more suitable than BPW, which represents the combination of this wearable system The machine learning model can be well applied to the community measurement data, and the wearable device combined with the above applied machine learning judgment also has a certain degree of analytical ability.
The device developed in this laboratory has been improved and combined with machine learning. It can not only be used in clinical analysis, but also in the community home environment. It has a certain degree of analytical ability and can provide more objective information in clinical applications. Assess, and at the same time allow more careful classification of the disease. Assist physicians to achieve the goals of early detection and early prevention in the community home environment.

１.社團法人台灣失智症協會.(2021). 失智照護原則http://tada2002.ehosting.com.tw/Support.Tada2002.org.tw/support_care04.html
2.台南市立台南醫院. (2021).醫療新聞
http://www.tmanh.org.tw/NewsInfo/NewsArticle?no=825
3. 衛生福利部.(2021).失智症診療手冊
4. 台灣家庭醫學醫學會(2019). 第35卷.第八期.失智與憂鬱
https://www.tafm.org.tw/ehc-tafm/s/index.htm
5.健康電子報.2015年12月97期. 【檢驗小百科】關於抽血這件「曉」事
https://epaper.ntuh.gov.tw/health/201512/health_5.html
6. 基醫醫訊2016年第一季第88期.
https://www.kln.mohw.gov.tw/public/medi_news/fd9b1ace892df592b9369804dd2e6a97.pdf
7. 台灣精神醫學會.2003第17卷第1期.
http://www.sop.org.tw/sop_journal/Upload_files/17_1/06.pdf
8. 陳均曄（2019）。非侵入式循環訊號量測系統應用於不同癌症之偵測與評估。國立臺灣科技大學醫學工程研究所碩士論文，台北市。
9. 陳冠彰（2016）。循環訊號量測應用於乳癌患者病程監控之研究。
10. 楊舒涵（2015）。都卜勒微循環血流與動脈血壓波形指標 應用於代謝症候群之研究。國立臺灣科技大學醫學工程研究所碩士論文，台北市。
11. 許家良（2014）。血壓波形的諧波分析應用於多囊性卵巢及中風之偵測及療效評估研究。國立臺灣科技大學應用科技研究所博士論文，台北市。
12. 劉人銘（2018）。循環量測應用於服用高血壓藥物患者之研究。國立臺灣科技大學醫學工程研究所碩士論文，台北市。
13. ladecola, C, Gottesman, R. F.(2018). Cerebrovascular Alterations in Alzheimer
Disease. Circulation Research, 123(4), 406-408
14. sklearn.svm.SVC — scikit-learn 0.24.2 documentation
https://scikit-learn.org/stable/modules/generated/sklearn.svm.SVC.html
15. sklearn.neural_network.MLPClassifier — scikit-learn 0.24.2 documentation
https://scikit-learn.org/stable/modules/generated/sklearn.neural_network.MLPClassifier.html
16. sklearn.naive_bayes.GaussianNB — scikit-learn 0.24.2 documentation
https://scikit-learn.org/stable/modules/generated/sklearn.naive_bayes.GaussianNB.html
17. sklearn.tree.DecisionTreeClassifier— scikit-learn 0.24.2 documentation
https://scikit-learn.org/stable/modules/generated/sklearn.tree.DecisionTreeClassifier.html
18. sklearn.ensemble.RandomForestClassifier— scikit-learn 0.24.2 documentation
https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html
19. sklearn.linear_model.LogisticRegression— scikit-learn 0.24.2 documentation
https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html
20. sklearn.discriminant_analysis. Linear Discriminant Analysis— scikit-learn 0.24.2 documentation
https://scikit-learn.org/stable/modules/generated/sklearn.discriminant_analysis.LinearDiscriminantAnalysis.html
21. sklearn.neighbors.KNeighborsClassifier— scikit-learn 0.24.2 documentation
https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.KNeighborsClassifier.html
22. 楊長仁（2020）。穿戴式脈波量測系統結合機器學習於社區民眾血管狀態評估。國立臺灣科技大學醫學工程研究所碩士論文，台北市。
23. Li, B., Gao, H., Li, X., Liu, Y., & Wang, M. (2006). Correlation between brachial-ankle pulse wave velocity and arterial compliance and cardiovascular risk factors in elderly patients with arteriosclerosis. Hypertension research : official journal of the Japanese Society of Hypertension, 29(5), 309–314. https://doi.org/10.1291/hypres.29.309