Predicting severe clinical events by learning about life-saving actions and outcomes using distant supervision

Area of Science:

• Medical informatics
• Clinical decision support
• Machine learning in healthcare

Background:

• Medical errors, particularly missed signs of patient deterioration, contribute significantly to mortality.
• The heterogeneity of physiological patterns in patients presents a challenge for early detection of acute organ failures.

Purpose of the Study:

• To implement and evaluate risk prediction models for acute onset diseases using gradient boosted trees.
• To utilize physiological variables and clinical outcome data for training predictive models.

Main Methods:

• Employed gradient boosted tree models trained on physiological variables from ICU admissions.
• Used outcome-related interventions and discharge diagnoses as labels for supervised learning.
• Validated models on two large datasets: MIMIC-3 and UW-CDR.

Main Results:

• Achieved highest F1 score of 0.6018 for predicting acute lung injury interventions within 24h (MIMIC-3).
• Median F1 score of 0.3850 across all acute organ failures in both datasets.
• Highest F1 score of 0.6301 for classifying acute lung injury status at discharge (MIMIC-3).

Conclusions:

• Supervised machine learning models trained with intervention and diagnosis data can effectively predict the risk of acute organ failures.
• This approach demonstrates potential for improving early identification and management of critical illnesses.