Machine Learning as a Novel Method to Support Therapeutic Drug Management and Precision Dosing

Abstract

Machine learning is now being introduced in medicine, with the idea that the vast multidimensional clinical data and experiences documented in electronic health records can assist in the care of future patients.1 The ability of a machine learning model to find statistical patterns across millions of features and examples is what enables superhuman performance.2 By mining the complex and heterogeneous data from electronic health records (referred to as "big data"), patients can benefit from previous collective experience leading to improvements in diagnostic procedures, outcome predictions, and individualized treatments. Also in laboratory medicine, artificial intelligence and machine learning are finding their application.3 To date, population pharmacokinetic (popPK) models have been used to describe drug exposure over time in patients. With popPK models it is also possible to identify sources of between-patient variability in drug exposure such as body weight and renal function. During the modeling process, models with different numbers of compartments, types of elimination, and variability are evaluated.4 When a model has been designed that adequately describes the data, it can prospectively be used to guide drug dosing in an individual patient with the aim to achieve prespecified target concentrations with maximum precision and accuracy.5 At present this is largely done with stand-alone software packages for PK-guided dosing that are commercially or freely available, but it is expected that such tools will become more integrated into the electronic health records.6 In 2005 an online system for Immunosuppressant Bayesian Dose Adjustment (ISBA) was launched by the Department of Pharmacology and Toxicology at the Centre Hospitalier Universitaire at Limoges, France.7 This service was made available at no cost to transplantation centers worldwide.8 The investigators developed and validated popPK models, maximum a posteriori Bayesian estimators, and sparse sampling strategies for abbreviated area under the curve (AUC) estimation to generate individual PK parameters for immunosuppressive drugs for a variety of solid organ transplantation scenarios.9 This now widely used service provides recommendations for dose adjustment for individual transplanted patients, based on drug concentrations (preferably several samples within the first few hours post dose) uploaded via a secure portal and reported back to the requesting center. In the background, 145 Bayesian estimators are operational for the different posttransplantation times, ages (adult/pediatric), types of transplantation, associated immunosuppressive drugs, and drug measurement methods (Pierre Marquet, personal communication). In this issue of the journal, Woillard et al. present an innovative study comparing their traditional method of performing therapeutic drug management for mycophenolic acid, based on model-informed Bayesian estimation with the novel methodology of a machine learning approach.10 The paper is a follow-up to a recent publication in the journal by the same group, on machine learning for the precision dosing of tacrolimus.11 In both studies they used XGBoost, which stands for "Extreme Gradient Boosting," an open-source software library with which large data sets are fit into a mathematical model, while maintaining high computational speed. Interestingly, the machine learning approach outperformed the PK model-informed method for the indications and times after transplant tested, and the authors are planning to implement the machine learning models as an expert system in the ISBA service. Mycophenolic acid PK is notoriously difficult to predict in part because of unexpected between-occasion variability in absorption and occurrence of enterohepatic recycling. PK consults therefore frequently require clinical pharmacology expertise and input, something that has also been recognized as a potential hurdle for large-scale dissemination of available software tools for PK-guided dosing.12 The fact that the Limoges machine learning algorithm could predict mycophenolic acid exposure with the lowest imprecision in these independent populations of kidney and heart transplant recipients is promising. Similar findings were reported for tacrolimus machine learning models indicating that accurate estimation of abbreviated AUC based on two or three blood concentrations in kidney, liver, or heart transplant patients is feasible and results in better predictive performance than that achieved with the current PK model-based approach. These case examples represent important evidence in support of next steps toward implementation of machine learning algorithms for dose optimization in transplantation. The findings are particularly relevant as it would allow integration of such machine learning algorithms in cloud-based platforms or apps that would be accessible to clinical providers worldwide. Another application of machine learning in clinical pharmacology was recently reported by Center for Drug Evaluation and Research (CDER) researchers at the US Food and Drug Administration (FDA) who investigated how artificial neural networks could be applied to the modeling of drug exposure and response.13 Methods that use artificial neural networks have been studied in the past as promising tools in the field of drug discovery and development but have not yet found broad application.14, 15 The recent paper by the CDER group may change this as these investigators described good predicting performance of a type of recurrent neural network with specific applicability in predicting sequential data.13 The machine learning model was able to accurately predict individual pharmacodynamic (PD) response even when subjects were treated with different dosing regimens than were used in the model development. A summary of this work posted on the FDA's Regulatory Science Research and Education website also includes the suggestion that when scientists with backgrounds in machine learning, mechanistic modeling, and pharmacometrics interact, they will encounter many overlapping concepts.16 Machine learning approaches and more traditional mechanistic models can be complementary and when combined may improve the efficiency of drug development while also optimizing treatments for individual patients.16 Other recent examples of promising studies documenting the applicability of machine learning and artificial intelligence include machine learning modeling of free text data and adverse drug reactions coding for automatic identification of adverse drug reactions from unstructured electronic health record data and by integrating machine learning approaches in preclinical cardiotoxicity risk assessment and contributing patient factors analysis.17, 18 Language processing and text mining software tools can support data collection from electronic health records and provide important insights into real-world drug treatment outcomes.19 CPT's sister journals are also publishing on machine learning, including two articles highlighted in this issue of CPT.20, 21 What can be expected for the near future? Will we see the development of more machine learning algorithms to support therapeutic drug management and model-informed precision dosing for other categories of drugs? Will machine learning take over from popPK? And will the role of the clinical pharmacologist become reduced or actually be enriched? For the immunosuppressive drugs, extensive therapeutic drug management is performed following organ transplantation, and large data sets containing drug concentrations, drug dosages, and relevant patient characteristics are typically available.22 Such large data sets are needed in order to develop the machine learning algorithms. In other fields the limited availability of such data sets may be prohibitive for the development of machine learning algorithms. Therapeutic drug management of antibacterial, antifungal, and antiviral drugs is increasingly performed, especially in critically ill adult patients, and it is likely that for this indication large data sets will be available soon.23 A major advantage of popPK models is that the biological plausibility of the models is considered. The basis under a popPK model is a mechanistic model that is physiologically interpretable, whereas machine learning just aims to find the best algorithmic model with the best performance and with the smallest prediction error. Another important shortcoming of machine learning is that it does not allow for simulations. For example, an in silico comparison between different dosing strategies is not possible as machine learning is not flexible and not able to handle data that deviate from the learning data set and clinical scenario(s) on which the algorithm is based. Therefore it is likely that the introduction of machine learning will not lead to the disappearance of popPK, quantitative systems pharmacology, and other pharmacometrics methods—on the contrary. This view was also expressed in two recent expert opinion papers indicating that PK/PD modeling and simulation and artificial Intelligence and machine learning methodologies should be viewed as complementary.24, 25 Integration of these techniques has great promise for our field with the potential to enrich the future of predictive pharmacometrics and quantitative systems pharmacology. Experimental data from randomized controlled trials and pharmacometric and quantitative systems pharmacology approaches can be used not only to improve machine learning methods but also to provide ways for validating them.26 The application of continuous learning to iteratively update and refine prior knowledge with new data as they become available will greatly benefit early clinical development of new drugs as well as personalized precision therapeutics.

Topics & Concepts