| 摘要: | 隨著工業化與都市化的快速發展，環境中暴露於多種氣體（如硫化氫、一氧化碳、
二氧化碳、氧氣與氨氣等）的風險日益增加，對人體健康構成潛在威脅。危害氣體的吸
入可能對人體呼吸系統與心血管系統造成急性或慢性影響，導致生命體徵出現異常，甚
而危及生命安全。為此，本研究旨在建構一套可即時預測氣體暴露下之生理變化與健康
風險的智慧模型系統，作為未來智慧醫療與工業安全預警的核心工具。
本研究蒐集在多種氣體暴露條件下之生理感測資料，經過資料正規化、特徵工程與
演算法優化處理後，分別建構線性回歸(LR)、支援向量機（SVM）、決策樹（DT）與多
層感知器（MLP）等子模型，進行心率等關鍵生命體徵的迴歸預測。進一步透過二階段
集成式機器學習架構整合各子模型優勢，提升整體預測表現。本研究初步使用實驗動物
資料進行模型訓練與驗證，所建立之系統具有良好的泛化能力與轉譯潛力，其模型能夠
擷取氣體濃度與暴露時間對生理變化的高度關聯性。實驗結果顯示，本模型在多種氣體
環境下均能有效預測出生命體徵的變化趨勢，並具備即時預警與健康風險分級之能力。
未來透過引入人類生理資料並對參數做適當轉換，本系統可望廣泛應用於臨床急重症管
理、工業健康監控及智慧城市中的公共安全預防。;With the rapid advancement of industrialization and urban development, the risk of human
exposure to various hazardous gases such as hydrogen sulfide, carbon monoxide, carbon
dioxide, oxygen, and ammonia has significantly increased, posing potential threats to human
health. Therefore, this study aims to develop an intelligent predictive model system that can
monitor physiological changes in real-time and assess health risks associated with gas exposure.
This system will be a core component for future healthcare and industrial safety warning
platforms. Physiological sensing data collected under multiple gas exposure conditions were
processed through normalization, feature engineering, and algorithm optimization. Four
regression sub-models, Linear Regression (LR), Support Vector Machine (SVM), Decision
Tree (DT), and Multi-Layer Perceptron (MLP), were constructed to predict key physiological
indicators, such as heart rate. These models were further integrated into a two-stage ensemble
machine learning architecture to enhance overall prediction performance. Although initial
model training and validation were conducted using experimental animal data, the developed
system demonstrated strong generalizability and potential for translating findings to human
applications by capturing the correlations between gas concentration, exposure duration, and
physiological variations.
Experimental results indicate that the proposed model effectively identifies trends in
physiological signals under various gas conditions and supports real-time warnings and health
risk stratification. With future incorporation of human physiological data and dynamic
parameter adaptation, the system is expected to be widely applicable to clinical emergency
management, occupational health monitoring, and public safety prevention in smart city
infrastructures. |
