Smart Dosage Optimization using Machine Learning and Biomedical Engineering for Personalized Drug Therapy: A Review

Abstract

Background: Personalized medicine is revolutionizing modern medicine by customizing medicine for the individual patient based on their genetic, physiological and clinical profile. The problem of choosing the right dose, i.e., the maximum benefit versus minimum toxicity, is still one of the most important in clinical pharmacology, and conventional "population based" dosing recommendations for drugs do not account for significant inter-patient variability.
Objective: This review investigates the architectures, algorithms, clinical application and open challenges of the field and proposes ways to integrate this with machine learning (ML) to facilitate intelligent, patient specific dosage optimization.
Methods: The literature was systematically reviewed to highlight the effects and recent developments of ML-driven dosing and the enabling of biomedical-sensing technologies, categorizing the evidence based on the system architecture, algorithmic paradigms, biomedical-engineering enablers and comparative clinical outcomes.
Findings: In the studies reviewed, several electronic health record-based, genomic-based, laboratory-based, and continuous physiological signal-based ML dosing systems have been found to outperform traditional dosing systems in terms of dosing accuracy, reduction of drug-related toxicity, and increased patient adherence. Supervised models can be used to guide interpretable pharmacokinetic/pharmacodynamic predictions, deep learning can extract high dimensional, time-series physiological data, and reinforcement learning can provide closed-loop dosing for chronic and evolving disease states. These systems rely on wearable biosensors, smart drug delivery devices and Internet of Medical Things (IoMT) platforms to provide the real-time data streams.
Conclusion: The combination of ML and real time biomedical monitoring provides a viable avenue to safer, more effective and more individual drug therapy. To make this happen at scale, work needs to be done on data heterogeneity, model interpretability, clinical validation, regulatory acceptance, and on the performance of the models in different patient cohorts.