Multi-objective semi-supervised clustering to identify health service patterns for injured patients

Hadi Akbarzadeh Khorshidi, Uwe Aickelin, Gholamreza Haffari, Behrooz Hassani-Mahmooei

TL;DR

Researchers have developed a multi-objective semi-supervised clustering framework that groups injured patients based on their initial health service usage while simultaneously predicting their long-term recovery costs. By balancing unsupervised pattern recognition with supervised outcome prediction, the model identifies high-risk patients early, allowing healthcare providers to intervene before costs spiral.

What Problem Does This Paper Address?

Transport accidents are a leading cause of global mortality and long-term disability. For insurance providers and health systems, the "holy grail" is identifying which patients will require extensive, high-cost recovery as early as possible.

However, early-stage data is often sparse. Traditional unsupervised clustering can group patients who use similar services (e.g., both got X-rays), but these groups might not correlate with actual health outcomes. Conversely, standard supervised learning (like regression) can predict costs but fails to identify the underlying "types" or patterns of service utilization that lead to those costs. This paper bridges that gap by creating a model that cares about both: how patients look (similarity) and where they are headed (outcome).

Methods

The authors propose a multi-objective optimization model for semi-supervised clustering. The "semi-supervised" aspect comes from the fact that the model optimizes two conflicting goals at once:

  1. Unsupervised Goal: Minimize the $k$-medians clustering cost function. This ensures that patients within a group have similar patterns of health service use.
  2. Supervised Goal: Minimize the cross-validated root-mean-square error (RMSE) of a regression model. This ensures the clusters are actually predictive of the total recovery cost.

To solve this, they employed an evolutionary computation approach. Instead of starting with one set of clusters, they start with a "pool" of potential solutions with varying numbers of clusters ($K$). They use stochastic gradient descent to update these solutions over 500 iterations.

Because you cannot perfectly satisfy both objectives simultaneously, the authors use Pareto optimal solutions—a set of "best compromise" solutions where you can't improve one objective without hurting the other. Finally, they use a Decision Tree to turn the complex cluster assignments into simple, interpretable "if-then" rules for clinicians.

Datasets

The study utilizes real-world administrative data from the Compensation Research Database in Victoria, Australia, provided by the Transport Accident Commission (TAC).

  • Scope: 20,692 unique patients with non-fatal injuries occurring between 2010 and 2012.
  • Data Points: Over 1.6 million payment records.
  • Features: 19 health service categories (e.g., ambulance, pharmacy, hospital, paramedical) used within the first week post-injury.
  • Outcome: Total cost of recovery.

Key Insights and Ideas

  • The Power of "And": The most significant insight is that optimizing for clustering and prediction together yields better results for both than optimizing for either alone. A model that only cares about cost (regression) creates "clusters" that are statistically messy, while a model that only cares about similarity ($k$-medians) creates groups that tell you nothing about the patient's financial future.
  • Automatic K-Selection: By including the number of clusters as a variable in their evolutionary pool, the algorithm naturally converged on an optimal number of groups (3) without the researchers needing to guess beforehand.
  • Early Warning Signs: The model revealed that imaging and therapeutic services in the first week are the most critical "canaries in the coal mine" for identifying high-cost recovery trajectories.

Key Results

  • Superior Partitioning: The multi-objective model achieved an average silhouette width of 0.505, significantly higher than the $k$-medians-only model (0.109) and the regression-only model (-0.019). This indicates much cleaner, more distinct patient groupings.
  • Predictive Accuracy: The model maintained a low cross-validation error (1.239), outperforming the $k$-medians-only approach (1.527) and remaining competitive with pure regression.
  • High Interpretability: The decision tree extracted from the clusters was remarkably accurate, with a misclassification rate of only 2.1%, far lower than standard classifiers like Linear Discriminant Analysis (13.8%) or Naïve Bayes (31.4%).

Limitations and Future Work

While the model is highly effective at predicting costs, the authors acknowledge that "cost" is only one measure of recovery. They suggest that future work should apply this multi-objective framework to other outcomes, such as Return to Work (RTW) status or clinical quality-of-life metrics. Additionally, while the model handles the first week of data well, exploring how these patterns evolve over longer timeframes could provide even deeper insights into patient trajectories.

Summary updated

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.
BibTeX
@inproceedings{khorshidi2019multi,
  author = {Hadi Akbarzadeh Khorshidi and Uwe Aickelin and Gholamreza Haffari and Behrooz Hassani-Mahmooei},
  title = {{Multi-objective semi-supervised clustering to identify health service patterns for injured patients}},
  booktitle = {Health Inf. Sci. Syst},
  year = {2019},
  url = {https://openreview.net/forum?id=gnaWCXYgrV},
}

Loading related content...

Figure