No More False Alert: Contrastive Learning for Predicting Health Deterioration from Imbalanced Care Records.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Healthcare Informatics

Background:

• Long-term care facilities monitor older adults' health using daily records.
• Predicting health deterioration aids timely medical intervention and care adjustments.
• Class imbalance in rare deterioration events leads to false positives and alert fatigue.

Purpose of the Study:

• To develop an outcome-based contrastive loss for imbalanced binary classification.
• To improve prediction of next-day health deterioration using care records and meteorological data.
• To reduce false positives and alert fatigue in health deterioration prediction systems.

Main Methods:

• Proposed an outcome-based contrastive loss function.
• Contrasted actual deteriorating samples against false alarms within mini-batches.
• Utilized care records and meteorological data for prediction.
• Employed pair-design ablations and UMAP for analysis.

Main Results:

• Improved precision from 3.97% to 12.94% compared to standard cross-entropy with oversampling.
• Limited F1-score decrease to 0.71 percentage points.
• Demonstrated clearer separation between true deterioration and false alarms in learned representations.

Conclusions:

• The proposed loss function effectively reduces unnecessary alerts in health deterioration prediction.
• This approach offers a viable direction for developing reliable alert systems in care settings.
• Reduced alert fatigue and improved detection support better patient outcomes and reduced healthcare burdens.