Human-in-the-loop Reinforcement Learning for personalized and adaptive interventions in wellness and healthcare: A systematic literature review.

BACKGROUND

Reinforcement Learning (RL) is emerging as a promising technique to support personalized and adaptive clinical interventions. This is thanks to its ability to model sequential decision making and to learn from longitudinal data. The deployment of RL raises challenges regarding interpretability and integration into clinical workflows. These concerns have motivated interest in human-in-the-loop approaches, in which individuals retain an actual or intended role before an RL-supported decision affects care. However, the use of RL in healthcare remains fragmented and poorly standardized. A systematic review of the literature is therefore timely, to identify current state, directions, challenges and research gaps.

METHODS

A systematic review was reported in accordance with PRISMA guidelines. ScienceDirect, Scopus, PubMed, IEEE Xplore, and Web of Science databases were searched for all records available up to 21 January 2026, encompassing studies applying RL to personalized interventions in healthcare and wellness fields and including the concept of "human-in-the-loop". Extracted information contained diseases, applications, RL-methods, dataset dimensions, integration with other techniques, and performance evaluation strategies.

RESULTS

Forty-one papers met the inclusion criteria and were examined in detail. RL was mainly found in the healthcare rather than in the wellness field, especially in type 2 diabetes and other chronic diseases, and intensive care settings, where the sequential nature of decision-making processes plays a significant role. Data structure and information vary extensively, creating significant comparability challenges. The methodologies show strong heterogeneity in algorithms and evaluation metrics, limiting reproducibility and generalizability. From the disease-focus point of view, cardiovascular applications are underrepresented, despite cardiovascular diseases being the leading cause of death globally and the potential value of adaptive decision-making in cardiovascular disease management.

CONCLUSIONS

RL offers an opportunity to personalize treatments and optimize clinical decision-making processes employing sequential data. However, the lack of shared methodological standards, transparency in model descriptions, and limited number of prospective studies constitute significant barriers to its adoption. In conclusion, better documented datasets, comparable metrics, and prospective evaluations are required to establish clinical effectiveness and support future implementation.