Adaptive Music Selection Using Artificial Intelligence as a Conceptual Approach to Dementia Care

Journal of Psychiatry Reform vol 13, #2. February 16, 2026

Elizabeth Bighiu¹, Ana Hategan, MD, FRCPC²

Author Information

1  Second Year Computer Science/AI Major Undergraduate Student pursuing a Bachelor of Computing (BCmpH) degree, Queen’s University, Kingston, Ontario, Canada; [email protected]

2  Clinical Professor, Geriatric Psychiatrist, Division of Geriatric Psychiatry, Department of Psychiatry and Behavioural Neurosciences, Michael G. DeGroote School of Medicine, Faculty of Health Sciences, McMaster University, Hamilton, Ontario, Canada. [email protected]. ORCID iD https://orcid.org/0000-0003-0221-1154

Music has long been used as a supportive, non-pharmacological approach to improve mood and reduce distress in older adults [1, 2]. In people living with cognitive impairment such as dementia (or a major neurocognitive disorder), agitation is a frequent and difficult symptom to manage, particularly in long-term care homes, and non-pharmacological interventions are often preferred [3]. While music interventions are commonly used, they usually involve fixed playlists that do not change in response to how an individual reacts over time [4].

Recent advances in artificial intelligence (AI) raise the possibility of personalizing music selection in a more adaptive way. Rather than relying on a pre-selected playlist, an AI-based system could adjust the type of music played based on observable responses, such as whether a song is listened to fully, skipped quickly, or appears to coincide with a calmer behavior. This idea goes beyond simply creating playlists and instead focuses on learning which music seems most helpful for a specific individual. This adaptive approach is motivated by the high degree of uncertainty inherent to music preferences: across a large and diverse set of songs, it is rarely obvious in advance which selections will be effective for a given individual.

To explore this idea, one of the authors (EB) developed an AI-driven music selection system tool, structured using a reinforcement learning framework. The problem is modeled as a Markov Decision Process (MDP) (i.e. a mathematical model for sequential decision making when outcomes are uncertain [5]), where each state represents a currently playing song and actions correspond to transitions between songs. The system learns a policy that determines which song to play next based on interactions, specifically whether a song is liked or skipped. These interactions provide feedback on how well a song aligns with an individual’s preferences. The goal is to minimize a long-term cost associated with unfavorable responses. Disliking a song incurs a higher cost, while liking it or indicating positive engagement incurs a lower cost. The cost at each step is a scalar value, and the total objective is to minimize the discounted sum of these costs over successive song selections. Rather than assuming the environment is fully known, the system learns from experience. A value function is defined to capture the expected discounted cost of future song selections, and policy gradient methods are used to iteratively improve the selection strategy. Here, the value function represents an estimate of how beneficial a song choice is expected to be in the long run, based on the sequence of future responses it is likely to produce. Policy gradient methods then adjust the song-selection strategy by increasing the likelihood of choice that is associated with a lower expected cost, using feedback gathered from repeated interactions over time. Because exact gradients are difficult to compute in practice, stochastic gradient descent is used to update the policy incrementally based on sampled interactions.

Choosing the appropriate population is important when discussing potential outcomes. Dementia represents a clear clinical population in which agitation is a recognized and ongoing management challenge. Individuals with dementia living in long-term care settings may be an appropriate group in which to explore adaptive music-based approaches. Memory loss in dementia often affects more recent memories first, while memories formed earlier in life may remain accessible for longer periods. Music from earlier life stages may therefore still be recognizable and emotionally meaningful. An adaptive system could potentially identify which songs or time periods are associated the most with calmer responses and prioritize those selections, rather than applying the same music to all individuals. This process may help reduce agitation and support emotional regulation.

Such a system would not replace caregivers or clinical judgment. Instead, it could function as a supportive tool to help tailor music interventions more consistently, particularly in care environments where staff time is limited. Any potential benefit would need to be evaluated carefully, as responses to music are highly individual and may change over time.

To test the feasibility of this approach, the system was trained using a dataset of 35 songs with diverse audio features and four simulated listener profiles representing different musical preferences. Simulated users were implemented at this stage to allow controlled, repeatable experimentation and to evaluate whether the system could learn adaptive behavior before involving human participants. While the initial dataset was limited in size, the number of songs could be easily increased, and curated to include music older adults would find preferable. Virtual users interacted with the system, allowing it to learn which features were associated with favorable responses for different listener types. When a new, previously unseen user was introduced, the system was able to rapidly infer their preference pattern based on limited interaction data and adjust its recommendations accordingly. This suggests that adaptive systems may be capable of personalizing music selection efficiently, even with sparse feedback.

This concept is speculative and would require empirical testing before any clinical application. Ethical issues, including privacy, consent, and appropriate oversight, would need to be addressed. Small pilot studies focusing on feasibility and acceptability, as well as caregiver-reported agitation or calmness, could help determine whether this approach is worth further investigation.

In summary, adaptive AI-guided music selection may represent a promising area for future research in dementia care. Given the low-risk nature of music-based interventions and the need for non-pharmacological strategies to manage agitation, further exploration of this idea may be warranted. If effective, this approach may help reduce agitation episodes, potentially improving quality of life for long-term care residents and easing caregiver burden.

Acknowledgement: We sincerely thank Prof. Bahman Gharesifard Ph.D, ([email protected]), Department of Mathematics and Statistics, Queen’s University, Kingston, Ontario, Canada for contributing to the theoretical research of this adaptive AI-guided music selection tool.

Funding: The author (EB) received financial support through a paid summer 2025 research internship at Queen’s University. The funder had no role in study design, analysis, or interpretation.

Conflicts of Interest: None declared.

References

1. Ma G, Ma X. Music intervention for older adults: Evidence map of systematic reviews. Medicine (Baltimore). 2023;102(48):e36016. doi: 10.1097/MD.0000000000036016. PMID: 38050267; PMCID: PMC10695625.
2. Vincenzi M, Borella E, Sella E, Lima CF, De Beni R, Schellenberg EG. Music listening, emotion, and cognition in older adults. Brain Sci. 2022;12(11):1567. doi: 10.3390/brainsci12111567. PMID: 36421891; PMCID: PMC9688894.
3. Jones E, Aigbogun MS, Pike J, Berry M, Houle CR, Husbands J. Agitation in dementia: Real-world impact and burden on patients and the healthcare system. J Alzheimers Dis. 2021;83(1):89-101. doi: 10.3233/JAD-210105. PMID: 34250934; PMCID: PMC8461728.
4. Sakurai K, Togo R, Ogawa T, Haseyama M. Controllable music playlist generation based on knowledge graph and reinforcement learning. Sensors (Basel). 2022;22(10):3722. doi: 10.3390/s22103722. PMID: 35632130; PMCID: PMC9144078.
5. Alagoz O, Hsu H, Schaefer AJ, Roberts MS. Markov decision processes: a tool for sequential decision making under uncertainty. Med Decis Making. 2010;30(4):474-83. doi: 10.1177/0272989X09353194. Epub 2009 Dec 31. PMID: 20044582; PMCID: PMC3060044.