| 摘要: | 生理訊號(“Physiological Signal” : are defined as multichannel readings from the central and autonomic nervous system that carry meaningful information in terms of actions, responses, feelings, and behavior. 取自 Computer Science Review, 2021)作為中樞與神經系統對於動作、反應、感知與行為所帶出具有意義的資訊表現，透過各式監測設備以時間序列取樣方式收集數據來源，提供不限於醫療、照護與醫學研究等行為所需的生理狀態探索與觀察。
近年來有別於傳統ECG/PPG/sphygmomanometer等的各類新穎監測手段推陳出新，尤其在非接觸式的監測應用發展甚受矚目，這類型的技術方案主要用意不在於取代傳統的醫材，更著重於一般生活領域與長期照護場所的運用，實現”連續”與”零束縛”的生理訊號收集目的，凸顯對”可能風險、早期發現”與”隨時隨地、以防萬一”的使用價值。然而此類的監測手段容易遭受外在環境與受測者自主活動的干擾，為能順利導入使用場域，一套合用的訊號演算法勢必為成敗的關鍵。然而不同的硬體感測方案具備不同的訊號特性，每每需要專門開發一套對應的演算法，本研究旨在探討是否有機會發展一套較為通用型的算法架構，可以既”簡易”又”相容”各種檢測方案，降低使用者開發門檻，隨著市場上硬體感測方案陸續在性能及成本上的優化，使用一套易上手又具分析意義的演算法，可加速產業界商品化時程與多元性，造福更多終端使用者。
對於生理訊號的監測與分析基本建立於訊號表現的”周期性”與”振幅度”上，本研究成果所採用的簡易數據轉換即忠實地表現了前述兩種特性，同時追求較大相容度，選用的測試方案為低頻2.4GHz的生理雷達產品，偵測內容為呼吸率與心跳率，其頻段干擾最多，但可容許安裝距離相較高頻段產品遠，達2公尺，隨帶有一些訊號基頻偏移情況，在此狀態下實驗結果呼吸演算平均誤差率可在2 rpm以內，心跳平均誤差率可於8 bpm以內，另可在相同數據下運用簡易公式換算，成功識別”擾動”、”有生理訊號”與”無生理訊號”三類狀態，提供臨床照護重要風險指標。
最後，利用前述的數據轉換，結合基本的類神經網路MLP架構，訓練兩種訊號型態包含”有生理訊號”與”無生理訊號”，亦能順利完成建模，測試集擁有最佳98%的分類器成效，未來有機會延伸應用在特定生理表徵的模型建立與危險族群的風險預警，證明本研究成果具備：一套方法、多元使用、簡易且相容度高的預期優勢。
;Physiological Signal: are defined as multichannel readings from the central and autonomic nervous system that carry meaningful information in terms of actions, responses, feelings, and behavior. (Taken from Computer Science Review, 2021). They are collected through various monitoring devices using time series sampling methods to provide data sources for the exploration and observation of physiological states required for activities such as medical care, caregiving, and medical research.
In recent years, various innovative monitoring methods, different from traditional ECG/PPG/sphygmomanometers, have emerged. Particularly, the development of non-contact monitoring applications has garnered significant attention. These technological solutions are not intended to replace traditional medical devices but rather emphasize their application in daily life and long-term care scenario. The goal is to achieve "continuous" and "zero-restriction" collection of physiological signals, highlighting their value for "potential risk, early detection," and "anytime, anywhere, just in case" use. However, these monitoring methods are susceptible to interference from the external environment and the autonomous activities of the elder people or patients. To successfully implement these methods in practical use, an appropriate signal algorithm is crucial. Different hardware sensing solutions have different signal characteristics, often requiring the development of specialized algorithms. This research aims to explore the possibility of developing a more generic algorithm framework that is both "simple" and "compatible" with various detection schemes, reducing the development barrier for users. As the performance and cost of hardware sensing solutions in the market continue to optimize, applying an easy-to-use and meaningful analytical algorithm can expedite the commercialization process and diversity in the industry, benefiting more end users.
The monitoring and analysis of physiological signals fundamentally relies on the "periodicity" and "amplitude" of the signal expression. The simple data transformation adopted in this study faithfully represents these two characteristics while pursuing greater compatibility. The selected test scheme involves a low-frequency 2.4GHz physiological radar product, detecting respiratory rate and heart rate. Although its frequency band is highly susceptible to interference, it allows for a greater installation distance compared to high-frequency products, reaching up to 2 meters. Despite some signal fundamental frequency offset situations, experimental results demonstrate that the average error rate for respiratory calculations can be within 2 rpm, and the average error rate for heart rate can be within 8 bpm. Additionally, using a simple conversion formula with the same data, it is possible to successfully identify three states: "disturbance," "presence of physiological signal," and "absence of physiological signal," providing crucial risk indicators for clinical care.
Finally, using the a forementioned data transformation combined with a basic neural network MLP (Multi-Layer Perceptron) architecture, modeling for two signal types, including "presence of physiological signal" and "absence of physiological signal," can be successfully completed. The test set achieved an optimal classifier performance of 98%, indicating potential future applications in establishing models for specific physiological features and risk warnings for vulnerable groups. This proves that the research results have the expected advantages of being a versatile, simple, and highly compatible method for multiple uses. |
