Abstract
This study develops a pattern recognition method that identifies patterns based on their similarity and their association with the outcome of interest. The practical purpose of developing this pattern recognition method is to group patients, who are injured in transport accidents, in the early stages post-injury. This grouping is based on distinctive patterns in health service use within the first week post-injury. The groups also provide predictive information towards the total cost of medication process. As a result, the group of patients who have undesirable outcomes are identified as early as possible based health service use patterns.
Introduction
Transport accidents are a leading cause of global mortality and long-term disability. For insurance providers and health systems, identifying patients who will require extensive, high-cost recovery early in the process is a significant challenge. Early-stage data is often sparse, and traditional methods struggle to balance identifying service utilization patterns with predicting actual health outcomes. This study develops a multi-objective semi-supervised clustering framework to bridge this gap, grouping patients based on initial health service usage while simultaneously predicting long-term recovery costs.
Methodology
The researchers utilized administrative data from the Compensation Research Database in Victoria, Australia, provided by the Transport Accident Commission (TAC). The dataset encompassed 20,692 unique patients with non-fatal injuries occurring between 2010 and 2012, involving over 1.6 million payment records. Features included 19 health service categories (e.g., ambulance, pharmacy, hospital, paramedical) used within the first week post-injury, with the total recovery cost as the primary outcome.
The proposed model employs a multi-objective optimization approach to balance two conflicting goals:
- Unsupervised Goal: Minimizing the $k$-medians clustering cost to ensure patients within a group have similar service usage patterns.
- Supervised Goal: Minimizing the cross-validated root-mean-square error (RMSE) of a regression model to ensure clusters are predictive of total recovery costs.
An evolutionary computation approach was used to navigate this trade-off, starting with a pool of potential solutions and evolving them over 500 iterations toward a Pareto optimal front—a set of "best compromise" solutions where one objective cannot be improved without degrading the other.
Results and Discussion
The multi-objective model successfully identified three distinct patient clusters. It achieved an average silhouette width of 0.505, indicating significantly cleaner partitioning than single-objective models (k-medians only: 0.109; regression only: -0.019). Furthermore, the model maintained high predictive accuracy with a low cross-validation error of 1.239, outperforming the k-medians-only approach.
A key finding was the high interpretability of the resulting clusters. By applying a decision tree to the cluster assignments, the researchers derived simple "if-then" rules that clinicians can use. The decision tree revealed that imaging and therapeutic service costs in the first week are the most critical predictors of high-cost recovery trajectories.
The decision tree achieved a remarkably low misclassification rate of 2.1%, significantly outperforming standard classifiers like Linear Discriminant Analysis (13.8%) and Naïve Bayes (31.4%). This suggests that the identified clusters are not only statistically sound but also practically applicable in a clinical setting.
Conclusion
This study demonstrates that optimizing for both clustering and prediction simultaneously yields superior results for identifying health service patterns. The framework provides an early warning system for high-risk patients, allowing for timely intervention before costs escalate. Future work could extend this multi-objective approach to other outcomes, such as Return to Work (RTW) status or clinical quality-of-life metrics, and explore how these patterns evolve over longer timeframes.
Summary updated