Reducing Artifact Preprocessing in Heart Rate Variability-Based Personalized Psychosis Prediction Using Adaptive Long...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Schizophrenia spectrum disorders are associated with physiological changes.
• Heart Rate Variability (HRV) data from wearables offers a potential biomarker for monitoring these changes.
• Previous psychosis prediction models often rely on extensive data preprocessing, including artifact removal.

Purpose of the Study:

• To investigate the efficacy of Long-Short-Term Memory (LSTM) networks for psychosis prediction using unprocessed Heart Rate Variability (HRV) data.
• To determine if artifact removal preprocessing is necessary for accurate personalized relapse prediction in schizophrenia spectrum patients.
• To evaluate the impact of omitting data cleaning steps on prediction accuracy.

Main Methods:

• Collected sleep HRV recordings from 7 patients (7-113 days each) and validated on a 30-patient psychosis cohort.
• Computed HRV characteristics directly from unprocessed time series, bypassing artifact correction.
• Utilized LSTM networks with patient-specific sequence lengths to learn temporal relationships and individual physiological trends in HRV.

Main Results:

• The LSTM model achieved a mean F1 score of 0.9817 on the initial 7-patient cohort.
• Omitting artifact-removal preprocessing did not decrease, and in some cases improved, psychosis prediction accuracy.
• The personalized prediction model demonstrated strength across diverse patient profiles and clinical settings.

Conclusions:

• Personalized psychosis prediction using wearable HRV data is feasible with LSTM networks.
• Artifact removal preprocessing may not be essential for accurate psychosis prediction from wearable HRV data.
• This simplified approach holds promise for real-time monitoring and early intervention in schizophrenia spectrum disorders.