Automatic patient functionality assessment from multimodal data using deep learning techniques - Development and...

Area of Science:

• Digital Health
• Artificial Intelligence in Healthcare
• Biomedical Signal Processing

Background:

• Wearable devices and mobile sensors generate extensive real-time physiological and behavioral data.
• This data offers potential for remote patient monitoring, surpassing traditional methods.
• Challenges include data noise and missing observations, hindering robust model development.

Purpose of the Study:

• To develop an AI pipeline for predicting mobility impairment using digital biomarkers from wearable sensors.
• To address data challenges like missing observations and leverage unlabeled data through transfer learning.
• To evaluate the model's performance and interpretability in real-world datasets.

Main Methods:

• An attention-based Long Short-Term Memory (LSTM) neural network pipeline was proposed.
• Hidden Markov models were utilized to handle missing data.
• Transfer learning was employed to incorporate information from unlabeled samples.
• The approach was validated on two real-world wearable/mobile sensor datasets.

Main Results:

• The proposed pipeline outperformed a prior baseline in predicting mobility impairment (WHODAS 2.0).
• Attention heatmaps provided interpretability for the model's predictions.
• Task transfer learning enabled accurate prediction of generalized anxiety severity (GAD-7) using a smaller cohort.

Conclusions:

• The developed AI pipeline effectively predicts health outcomes like mobility impairment and anxiety from wearable sensor data.
• The model demonstrates robustness in handling noisy and incomplete data, crucial for remote patient monitoring.
• The approach offers interpretable insights and shows potential for broad application in digital health.