Risk perception modeling based on physiological and emotional responses / Ding Huizhe

Ding , Huizhe (2024) Risk perception modeling based on physiological and emotional responses / Ding Huizhe. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (154Kb)
	PDF (Thesis PhD) Download (4Mb)

Abstract

Risk perception refers to how individuals perceive objective risks. Although it initially emerged in social sciences, it has become a crucial aspect of safety science due to its significance in understanding unsafe behaviors. It can help safety managers develop a comprehensive understanding of risk based on traditional engineering risk assessment principles, facilitating the transition from the Safety I to Safety II paradigm. Therefore, accurate risk perception has become vital. This research aims to develop models for objectively assessing perceived risk. Previous studies have employed machine learning techniques to classify high and low-risk situations based on physiological responses. However, the performance of these algorithms in situations with closely comparable risk magnitudes remains uncertain. This issue is crucial as it directly impacts their practicality and generalization. To address this concern, four driving clips were selected as stimuli, including relatively low (1.87) and high (3.97) risk levels, as well as two clips with slight variations in their degree of riskiness (2.45 and 2.85, respectively). Fifty-five subjects were recruited to synchronously measure their physiological signals, including Electrodermal Activity (EDA), Heart Rate Variability (HRV), Pupil Diameter (PD), and Skin Temperature (ST). A Pleasure-Arousal-Dominance (PAD) model was used to induced and expressed mixed emotions. Subsequently, statistical analyses were performed to identify indicators that showed significant differences. These results varied significantly, including three emotional dimensions, two skin conductance indicators (EDR and EDL), and several ECG indicators (such as HF, LF/HF and A++) reflecting short-time changes. As the perceived risk level increased, subjects’ emotions experienced more negative, arousal and a diminished sense of control. In terms of physiological changes, there was an increase in sympathetic activity and a concurrent decline in the vagus nerve at a macro-level. However, the changes that resulted from consecutive heartbeats were characterized by rapid and erratic variations at a micro-level. Additionally, these observed significant differences were primarily attributed to variations in risk levels, rather than personal differences. In terms of feature importance, physiological and emotional indicators that showed significant differences or greater fluctuations demonstrated greater sensitivity. Finally, three base models, Artificial Neural Network (ANN), Random Forest (RF), Support Vector Classification (SVC), and two integrated models were trained to classify perceived risk using higher sensitivity features. The ANN demonstrated superior ability in distinguishing low and high-risk levels. However, when risk degrees were closely matched, the integrated model with weight adjustments based on base models outperformed ANN. To validate the research findings, a second experiment was conducted in a construction scenario, still utilizing two clips with closely matched risk degrees. It was demonstrated that the primary results derived from statistical analysis and machine learning modelling were remarkably consistent, thereby confirming the effectiveness and generalization of the proposed weight adjustment algorithm, particularly in situations with closely matched risk levels.

Actions (For repository staff only : Login required)