Machine learning approach to identify malaria risk in travelers using real-world evidence

Abstract

Background: Pre-travel consultation and chemoprophylaxis measures for malaria are a key component in the prevention of imported malaria in travelers. In this study we report a predictive tool for assessing personalized malaria risk in travelers based on the analysis of electronic medical records from travel consultations. The tool aims to guide physicians in the recommendation of appropriate prophylaxis prior to their trip. We also provide best-practice recommendations for pre-processing noisy and highly sparse real world evidence data. Methods: We leveraged a large EMR dataset, containing demographic information about travelers and their destination. The data has been previously preprocessed using various strategies to handle missing and unbalanced data. We compared multiple machine learning approaches to assess the risk of malaria acquisition in travelers during their travels. Additionally, a feature importance analysis was performed using SHAP (SHapley Additive Explanations) values to identify patterns associated with malaria risk. Results: Our study revealed that our XGB models achieved high predictive capacity (AUC >0.80). The most significant features predicting malaria infection during travel included travel destinations with low malaria risk, vaccination history, number of countries visited, age, and trip duration. Remarkably, we were able to obtain a reduced model with only five features. When comparing this model with a population of travelers recommended for malaria chemoprophylaxis, we observed that it was deemed necessary in only 40% of these travelers. This suggests that 60% received chemoprophylaxis despite having a low personalized risk of malaria. Conclusion: We have developed an algorithmic tool that utilizes a concise survey to generate a personalized travel risk assessment, effectively minimizing the prescription of unnecessary malaria chemoprophylaxis. Through the identification of patterns linked to predictions, our model significantly enhances the efficacy of pre-travel consultations.