The application of machine learning models in a resource-constrained environment

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Critical Care Medicine

Background:

• Machine learning models (MLMs) typically require large datasets, posing challenges for resource-constrained intensive care units (ICUs).
• Limited research exists on applying standard MLMs to small ICU datasets in low-resource settings.

Purpose of the Study:

• To evaluate the efficacy of MLMs in predicting mortality using a small ICU dataset from a resource-constrained institution.
• To assess the performance of a resource-constrained scoring system (TropICS) within MLMs.

Main Methods:

• MLMs (XGBoost, KNN) were applied to a prospective cohort of mechanically ventilated patients in rural Kenya.
• The Tropical Intensive Care Score (TropICS) was utilized as a key characteristic.
• Area Under the Receiver Operating Characteristic curve (AUC) was calculated to predict mortality.

Main Results:

• The study included 294 patients with a 60.2% mortality rate.
• ML models demonstrated strong predictive performance for mortality (XGBoost AUC = 0.82).
• The TropICS score showed performance comparable to APACHE-II and SAPS within MLMs.

Conclusions:

• MLMs can be effectively implemented in small ICU datasets within resource-constrained environments.
• ML models require validation before integration into clinical practice.
• Contextualized scoring systems like TropICS perform well within MLMs.