PT - JOURNAL ARTICLE AU - Andrew E. Levy AU - Minakshi Biswas AU - Rachel Weber AU - Khaldoun Tarakji AU - Mina Chung AU - Peter A. Noseworthy AU - Christopher Newton-Cheh AU - Michael A. Rosenberg TI - Applications of Machine Learning in Decision Analysis for Dose Management for Dofetilide AID - 10.1101/531285 DP - 2019 Jan 01 TA - bioRxiv PG - 531285 4099 - http://biorxiv.org/content/early/2019/01/27/531285.short 4100 - http://biorxiv.org/content/early/2019/01/27/531285.full AB - Initiation of the antiarrhythmic medication dofetilide requires an FDA-mandated 3 days of telemetry monitoring due to heightened risk of toxicity within this time period. Although a recommended dose management algorithm for dofetilide exists, there is a range of real-world approaches to dosing the medication. In this multicenter investigation, we examined the decision process for dose adjustment of dofetilide during the observation period using machine-learning approaches, including supervised, unsupervised, and reinforcement learning applications. Logistic regression approaches identified any dose-adjustment as a strong negative predictor of successful loading (i.e., discharged on dofetilide) of the medication (OR 0.19, 95%CI 0.12 – 0.31, p < 0.001 for discharge on dofetilide), indicating that these adjustments are strong determinants of whether patients “tolerate” the medication. Using multiple supervised approaches, including regularized logistic regression, random forest, boosted gradient decision trees, and neural networks, we were unable to identify any model that predicted dose adjustments better than a naïve approach. A reinforcement-learning algorithm, in contrast, predicted which patient characteristics and dosing decisions that resulted in the lowest risk of failure to be discharged on the medication. Future studies could apply this algorithm prospectively to examine improvement over standard approaches.