Suggestions for the Ethical Design of a Health Care Treatment Recommender System
Advances in modern medicine have enabled health care to become increasingly precise. This holds promise for patients, however, the increase in the number of possible treatment options can make it difficult for physicians to decide on the best for a particular patient. Developing a treatment recommender system could assist physicians in this task, but due to the high-stakes nature of health care, it is imperative that medical ethics be integrated into its design. After an overview of medical ethics, in this paper, three key design features are investigated – (1) the selection of reinforcement as a learning algorithm, (2) the exploration / exploitation ratio and reward function design, and (3) how recommendations are presented and whether patients should be allowed to opt out of suggestions. Ultimately, ethical design suggests the development of a system that uses batch reinforcement learning with some exploration, a reward function that seeks to maximize a compound quality measure, and allows for patients to opt-out of suggestions. Possible future work could include the development of a treatment recommender system in a lower-risk health care domain (such as in primary care) and optimizing the design before moving to higher-risk domains such as oncology. Due to the incredibly complex nature of health care, treatment recommender systems are likely many years from being completely functional and because of the risks inherent in treating patients, and even if implemented, will require physician oversight for the foreseeable future.
- Health Care
- Treatment Recommender System
- Reinforcement Learning
- Ethics