Machine Learning for Precision Medicine and Health Economics

Author: Meimei Chen¹, Martin Hoyle², Yuanyuan Gu³ and Noemi Kreif⁴

1. PhD student, Macquarie University Centre for the Health Economy (MUCHE), Australian Institute of Health Innovation, Macquarie University, New South Wales, Australia.
2. Professor of Health Innovation and Evaluation, Macquarie University Centre for the Health Economy (MUCHE), Australian Institute of Health Innovation, Macquarie University, New South Wales, Australia.
3. Senior Research Fellow, Macquarie University Centre for the Health Economy (MUCHE), Australian Institute of Health Innovation, Macquarie University, New South Wales, Australia.
4. Senior Research Fellow, Centre for Health Economics (CHE), University of York, York, United Kingdom.

The following is a summary of a research project by the authors. For more information, please contact Meimei Chen at Macquarie University Centre for the Health Economy (MUCHE) (email: meimei.chen@mq.edu.au).

With recent advances in genomics, computing, connectivity, and artificial intelligence, health innovation is moving towards precision medicine, which uses a person's genetic, lifestyle, and other information to select the treatments that are most effective for them. But how can we better predict the most effective treatment on an individual basis? A recent study by the Macquarie University Centre for the Health Economy (MUCHE) and the University of York Centre for Health Economics developed a new machine learning model that classifies patients into subgroups based on their characteristics, each associated with different treatment outcomes. Results show the new model predict patients’ outcomes more accurately compared to several other benchmark methods.

Typically, the data used to perform the analysis are collected from clinical trials, registries and electronic devices and include demographics (e.g., age, gender, medical history) and gene expression profiles. Sometimes the number of variables in the collected data is larger than the number of participants; for example, genetic data may contain tens of thousands of variables, and such a dataset is referred to as a high-dimensional dataset. In this study, four open-source, high-dimensional cancer datasets were used to train the new machine learning model. Then the trained model was used to predict the survival time of the new patients based on each individuals’ information. Based on the predicted results, patients with longer survival times were selected to form a subgroup of patients with enhanced treatment effects. The subgroup of patients selected by the new model had the longest median survival times compared to other traditional statistical and machine learning methods. Results suggest the new model outperforms other methods in identifying patients who will experience longer survival following treatment compared to other methods. This new model might help clinicians select the most appropriate treatment for patients, and pharmaceutical companies develop new treatments. For example, if companies are interested in finding a subgroup of patients with enhanced treatment effects after a failed phase III trial.

The new model might help Health Technology Assessment (HTA) agencies better evaluate precision medicines. The goal of HTA is to provide policymakers with the necessary information to better understand the benefits of health technologies and make better funding decisions. In precision medicine, health technologies such as drugs are designed for specific subgroups of patients with better outcomes, rather than for entire patient populations. Thus, HTA evaluates new health technologies for precision medicine in specific patient subgroups rather than in the population as a whole. Compared to several other benchmark methods, the new model can better identify specific patient subgroups with enhanced effects. Therefore, the new model can potentially be applied to HTA to provide policymakers with more accurate information about precision medicine.

With precision medicine, healthcare has the potential to shift from expensive long-term management of chronic diseases to early intervention, which requires accurate identification of high-risk patients across the population. In addition, by accurately giving patients the most effective treatment, precision medicine can save the cost of unnecessary tests and procedures. Thus, the new model in this study has the potential to improve the identification process, patients’ health, and healthcare resource use.