Development of an AI-enabled predictive model to identify the 'sick child' at a pediatric telemedicine and medication delivery service in Haiti.

Background: One of the most difficult challenges in pediatric telemedicine is to accurately discriminate between the 'sick' and 'not sick' child, especially in resource-limited settings. Models that flag potentially 'sick' cases for additional safety checks represent an opportunity for telemedicine to reach its potential. However, there are critical knowledge gaps on how to develop such models and integrate them into electronic clinical decision support (eCDS) tools. Methods: To address this challenge, we developed a study design that utilized data from paired virtual and in-person exams at a telemedicine and medication delivery service (TMDS) in Haiti. Providers were allowed to mark respondent data as potentially unreliable. Artificial intelligence /machine learning (XGBoost) was applied to analyze paired data from participants across three consecutive implementation studies. Model derivation focused on identifying 'sick' patients (not-mild) and those requiring escalation. An ensemble method, based on gradient boosted decision-trees, was used given the limited sample size. The area under the receiver operating characteristic curve (AUC) was the primary outcome measure. Results: A total of 683 paired records were available for this secondary analysis from 2225 participants enrolled. The median age was 15 months and 47% were female. For prediction of a 'sick' child, we found an AUC of 0.82 (95% CI 0.78-0.86) after 5-fold cross validation; calibration slope and intercept were 1.31 (95%CI:1.09-1.53) and 0.04 (95%CI:-0.14-0.23), respectively. For prediction of escalation, we found an AUC of 0.78 (95%CI:0.74-0.81); calibration slope and intercept were 0.63 (95%CI:0.52-0.74) and 0.05 (95%CI:0.52-0.74), respectively. Accounting for data marked as potentially unreliable had mixed effects. Interpretation: These methods and findings offer an innovative and important proof-of-concept to improve pediatric telemedicine. The models require external validation prior to eCDS integration and deployment. Once validated, the models are designed to provide a critical safety check for experienced providers and digitally convey expertise to new providers.