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

TL;DR

The paper introduces Polygrid, an interpretable recommendation model tailored to psychometric data from gerontological primary care. By mapping multidimensional assessment scores to polygons in a unit disc and learning label-specific weights, Polygrid provides faithful, visual explanations that clinicians can inspect, enabling expert-in-the-loop decision making. Offline evaluations show Polygrid is competitive on multiclass and multilabel tasks, with some limitations in label ranking, while a dedicated user study demonstrates that the diagrammatic explanations support interpretation and speed for simple tasks. The work advances practical, interpretable recommender systems for dense, structured healthcare data and outlines clear directions for generalization, data sharing, and adaptation to diverse care settings. Its approach promises to reduce variability in referrals and improve alignment between recommendations and standardized CGA practices.

Abstract

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.

Figures (16)

• Figure 1: An example of the proposed explanation style. Legend: The explanation diagram is composed of a number of radar charts disposed on a rectangular grid. There are three types of charts: (a) assignment charts, which are placed on the first row, (b) assessment charts, on the first column, and (c) the matching charts.
• Figure 2: A structural equation model represented by a path diagram. Note: Figure \ref{['fig:background:selvm:cfa2']} shows an adaptation of the path diagram for the second order confirmatory factor analysis (CFA) model from bollen1989semlv. This model has been employed in prominent models used in gerontological research, including the instruments described in the Annex \ref{['section:background:cga:instruments']}. In this diagram, the latent variable model is composed of the subgraph induced by the latent variables, namely $\eta, x_1 \ldots x_3$. It encodes the fact that the second order variable $\eta$ causes variations in the first order variables $x_1 \ldots x_3$, but the latter do not influence each other. There are three measurement models, one for each of the first order variables (i.e., $x_1$ to $x_3$). A measurement model is composed of the subgraph induced by a given latent variable and its corresponding indicators. Figure \ref{['fig:background:selvm:mm']} focus on the collapsed measurement model for the latent variable $\eta$ to show that its path diagram is a pictorial representation of a system of structural equations, and that the hypothetical covariance matrix $\Sigma$ is derived from this system.
• Figure 3: Results obtained from the execution of the review protocol, by stage
• Figure 4: The learning and prediction pipelines of the Polygrid model. Legend: The pipelines are depicted at the top, as two directed paths. Each node corresponds to an algorithm (in the text). The steps of the learning pipeline for multilabel classification are indicated by the nodes 1 to 5. The learning pipeline for label ranking uses the "shadow" nodes: 1, 7, 3, 4, and 8. The prediction pipeline is encoded similarly. The matrices depicted at the bottom of the diagram illustrate how the input data are transformed along the pipelines, e.g., assessments in $X$ are transformed into polygons in $Z$ by the algorithm corresponding to the step 1.
• Figure 5: An explanation diagram is composed of a number of radar charts disposed on a rectangular grid. The diagram being displayed has been annotated with overlay elements (in grey) to single out its constitutive elements. There are three types of charts: (a) the assessment charts, which are placed on the first column, (b) the assignment charts, on the first row, and (c) the matching charts. The assessment chart depicts the results of a patient assessment. The vertices of an assessment polygon represent the scores the patient obtained for each domain of the instrument. Each assessment chart has a tag, which informs the area of its corresponding polygon. Typically, an explanation diagram displays the assessment of a single individual, but this example includes an extra one to benefit our discussion. The assignment chart depicts the representation of a label in the feature space: the colours in the background correspond to the weights ascribed to different cells of the unit disc. The polygon in an assignment chart, whose vertices are the mean scores of all assessments assigned to its respective label, is called the class prototype. Each assignment chart has a tag, which informs the threshold value for its respective class. In multiclass assignments, this value is omitted because each case is assigned to a single label, as described next. In a matching chart, the polygon is a copy of the assessment polygon on the same row, except that it is filled with "the colours" of the assignment chart in the same column. Each matching chart has a tag, which informs the weighted area of its polygon. A green tag indicates that this value is greater than its respective threshold, and the case is assigned to the respective label.