Evaluating the impact of human expertise in human-centered AI: A case study on finger-tapping video analysis for dementia detection.

PURPOSE: Human-centered artificial intelligence (AI) plays a crucial role in medical research. This paper evaluates the impact of human expertise in AI systems, using dementia prediction as a case study. Specifically, plasma phospho-tau181 (ptau181) is employed as the ground truth for Alzheimer's disease (AD) to advance early detection and treatment strategies.

METHODS

In this empirical study, we investigated three distinct cases to explore AI's role in predicting ptau181 levels through finger-tapping analysis. Case 1 employed explicit features from finger movements combined with a Ridge regression model, emphasizing the interpretability enabled by human-engineered features. Case 2 introduced temporal dynamics using Long Short-Term Memory (LSTM) networks with displacement-vs-time data, highlighting how human experts can integrate temporal dependencies into AI models. Case 3 utilized a 3D Convolutional Neural Network (CNN) to autonomously extract temporal and spatial features from processed video data, showcasing AI's ability to learn and adapt with minimal human intervention.

RESULTS

Evaluation demonstrated Case 3's superior performance, illustrating the trade-offs between model effectiveness and human involvement. Case 1 provided interpretability through explicit feature engineering, while Case 2 introduced complexity with temporal dependencies.

CONCLUSION

This study underscores the crucial role of human-centered AI in medical research, particularly in predictive modeling for AD. By combining human expertise with advanced AI capabilities, we can unlock new avenues for early diagnosis and intervention, thereby advancing our understanding and treatment of complex AD conditions to improve patient outcomes.