The Application Of Machine Learning Models In Resource-Constrained Environments
*Jaewook Shin Doctor of Medicine1, *Kemnunto Otoki MD2, *Addison M. Heffernan 1, *Robert K. Parker MD, FACS2, Daithi S. Heffernan MD, FACS1
1Surgical Research, Rhode Island Hospital/Alpert Medical School of Brown University, Providence, RI; 2Department of Surgery, Tenwek Hospital, Bomet, Kenya
Background: Machine Learning Models(MLMs) which increasingly influence surgical decision-making often require large and complex datasets upon which to train. However, there is a paucity of literature pertaining to the ability to apply standard MLMs to small ICU datasets within resource-constrained institutions. Study Design: A prospective cohort of critically ill, mechanically ventilated patients from a teaching hospital in rural Kenya. Characteristics included ICU scoring systems designed for resource-constrained environments, including Rwanda-Mortality Predictive Model(RMPM) and Tropical Intensive Care Score(TropICS). Outputs included AUC of the ROC and feature importance table. Python based MLMs included XGBoost(XGB) and RandomForest(RF). AUC of ROC was calculated to predict mortality as the primary endpoint. Results: There were 325 patients, average age 37 years, 64% male, 23% trauma and overall mortality of 60.7%. With respect to mortality, there was no difference in male gender(65.7% versus 62.2%;p=0.52) or transfer from outside facilities(35.4% versus 31.4%;p=0.54). However, patients who died were more likely to have infection(44.3% versus 14.9%;p<0.001), acidosis(7.26 versus 7.34;p<0.001), renal failure(21.1% versus 7.8%;p<0.001), hemodynamic instability(30.3% versus 9.5%;p<0.001) and higher APACHE-II, qSOFA, RMPM, and TropICS scores. In predicting mortality, MLMs performed extremely well(AUC XGB=0.85 and RF=0.9). Within MLM feature importance, RMPM and TropICS performed similar to APACHE-II and SAPS. When including only regional scoring models (RMPM, TropICS) and excluding APACHEII, MLMs still performed well(AUC XGB=0.77). The highest feature importance determinants included physiology and regional scoring models. Conclusion: MLMs can be effectively applied to relatively small ICU datasets within resource-constrained environments. Further, regional ICU scoring systems(RMPM and TropICS) perform well within these MLMs.
Back to 2024 Abstracts