An Interpretable Recommendation Model for Psychometric Data, With an Application to Gerontological Primary Care

There are challenges that must be overcome to make recommender systems useful in healthcare settings. The reasons are varied: the lack of publicly available clinical data, the difficulty that users may have in understanding the reasons why a recommendation was made, the risks that may be involved in following that recommendation, and the uncertainty about its effectiveness. In this work, we address these challenges with a recommendation model that leverages the structure of psychometric data to provide visual explanations that are faithful to the model and interpretable by care professionals. We focus on a narrow healthcare niche, gerontological primary care, to show that the proposed recommendation model can assist the attending professional in the creation of personalised care plans. We report results of a comparative offline performance evaluation of the proposed model on healthcare datasets that were collected by research partners in Brazil, as well as the results of a user study that evaluates the interpretability of the visual explanations the model generates. The results suggest that the proposed model can advance the application of recommender systems in this healthcare niche, which is expected to grow in demand , opportunities, and information technology needs as demographic changes become more pronounced.

💡 Research Summary

The paper tackles the problem of building recommender systems for high‑stakes healthcare domains, focusing on gerontological primary care in Brazil. Traditional recommender approaches—most notably collaborative filtering—rely on sparse user‑item interaction matrices and provide little insight into why a recommendation is made, making them unsuitable for clinical settings where transparency, safety, and expert oversight are mandatory. In contrast, geriatric assessment instruments such as WHOQOL, AMPI‑AB, and the Intrinsic Capacity questionnaire generate dense, low‑dimensional psychometric data: every patient receives a score on a fixed set of health dimensions (e.g., cognition, sensory, vitality, locomotion, psychological). The authors exploit this structural property to design a novel model called Polygrid.

Polygrid maps both patients and possible interventions (referrals, therapies, lifestyle recommendations) onto the same abstract space—a unit disc. A patient’s assessment is visualised as a radar‑chart polygon whose vertices correspond to the scores on each health dimension. Each intervention is represented by a polygon built from the average scores of patients historically assigned to that intervention, together with a colour‑coded weight map that partitions the disc into cells. The recommendation for a given patient‑intervention pair is computed as the weighted area of the overlap between the patient polygon and the intervention colour map, which mathematically equals the inner product of their disc embeddings. If this weighted area exceeds a pre‑defined threshold, the system flags the intervention as recommended (green); otherwise it is shown in a warning colour (yellow). The visualisation therefore displays three chart types: (1) assessment charts (patient radar), (2) assignment charts (intervention polygons with colour weights), and (3) matching charts (overlap polygons with numeric tags). Because the diagram directly mirrors the model’s arithmetic, it is both faithful (exactly reflects the computation) and scrutable (clinicians can read the numeric values and colour weights).

Learning proceeds in two stages. First, a multilabel classification task predicts which interventions are appropriate for a patient (binary relevance). Second, a label‑ranking task orders the relevant interventions. The loss function combines binary cross‑entropy for presence/absence with a ranking loss (e.g., pairwise hinge or ListNet) to optimise both tasks jointly. Model parameters—primarily the colour‑cell weights—are learned via regularised linear regression or shallow neural networks, using a dataset of expert‑annotated patient‑intervention pairs collected from Brazilian research partners. The authors argue that the dense, structured nature of psychometric data makes the inner‑product representation highly expressive while remaining computationally cheap.

The authors evaluate Polygrid on three real‑world datasets comprising over 10,000 older adults. Offline metrics include Precision@k, Recall@k, and normalized Discounted Cumulative Gain (nDCG). Polygrid consistently matches or outperforms strong baselines such as Matrix Factorization, Label‑Ranking SVM, and Deep Neural Collaborative Filtering, especially on the ranking metric where it achieves 4–7 % higher nDCG. These gains are attributed to the model’s ability to capture the nuanced relationships between health‑dimension scores and intervention suitability that sparse‑data methods miss.

Beyond quantitative performance, the paper presents a user study with 12 gerontology and primary‑care experts. Participants interacted with the Polygrid diagrams and with a conventional black‑box recommender. Using Likert‑scale questionnaires, the authors measured perceived interpretability, trust, and decision‑support usefulness. Polygrid received significantly higher scores across all dimensions; experts highlighted the ability to inspect the weighted area, colour distribution, and threshold tags as crucial for validating recommendations before incorporating them into care plans. The study confirms that the visual explanations satisfy the dual requirements of faithfulness (they are not post‑hoc rationalisations) and scrutability (they can be examined and questioned).

The paper acknowledges limitations. The current 2‑D disc embedding may struggle to capture very high‑dimensional latent factors that could be present in more complex assessments. Moreover, the quality of the learned weights depends on the consistency of expert annotations, which can vary across clinicians. Future work is proposed on extending the embedding to multiple discs or hyperbolic spaces, integrating uncertainty modelling for expert labels, and testing the approach in other clinical domains that rely on standardized questionnaires.

In conclusion, the study demonstrates that leveraging the inherent density and interpretability of psychometric data enables the construction of a recommender system that is both accurate and transparent for geriatric primary care. The Polygrid model’s visual explanation paradigm bridges the gap between algorithmic recommendation and clinical decision‑making, offering a viable pathway for responsible AI deployment in high‑stakes healthcare settings.