Patient Similarity Analytics for Explainable Clinical Risk Prediction

Hao Sen Andrew Fang, Ngiap Chuan Tan, Wei Ying Tan, Ronald Wihal Oei, Mong Li Lee, Wynne HsuResearch 2021 · DOI

Abstract

Background: Clinical Risk Prediction Models (CRPMs) estimate the probability of disease outcomes based on patient characteristics. Despite their utility, a lack of explainability and interpretability has limited their routine adoption in clinical practice. Methods: This study demonstrates a patient similarity analytics approach to develop an explainable CPRM using real-world data from 10,059 patients with type-2 diabetes, hypertension, and dyslipidaemia in a Singapore public primary care clinic. Results: The patient similarity model achieved an AUROC of 0.718, comparable to traditional machine learning models like Random Forest (0.764) and SVM (0.766). A prototype web application was developed to provide narrative-based evidence for clinical decision-making. Conclusions: A patient similarity approach is a feasible and generalizable method for developing locally relevant, interpretable CRPMs that facilitate shared decision-making.

Introduction

Clinical Risk Prediction Models (CRPMs) are powerful tools for estimating the likelihood of a patient developing a disease or complication. However, most state-of-the-art models act as "black boxes," providing a probability score without explaining the underlying reasoning. This lack of explainability (describing the model's process) and interpretability (the user's ability to understand the output) makes clinicians hesitant to trust these tools for high-stakes decisions. The authors aim to bridge this gap by using a similarity-based approach that mirrors natural clinical reasoning: comparing a current case to past experiences.

Methods

The researchers developed a CRPM focused on patients with "DHL" conditions—Type-2 Diabetes (D), Hypertension (H), and Dyslipidaemia (L). The study utilized a real-world dataset of de-identified Electronic Medical Records (EMR) from a public primary care polyclinic in Singapore, tracking 10,059 unique patients over a 10-year window (2009–2019).

Feature Selection and Weighting

The model utilized 69 discrete variables, including demographics, biomarkers, and medication history. To ensure clinical relevance, the researchers incorporated expert consensus to weight variables. For instance, "Duration of disease" was weighted heavily, while individual medication dosages were weighted lower.

Table 1: Key demographic and biomarker variables used to calculate patient similarity.

Similarity Metric and Outcomes

The core of the model is a weighted Euclidean distance calculation using a K-Nearest Neighbors (KNN) approach ($K=10$). Complications were defined using specific ICD-9 and ICD-10 codes across four main categories: Eye, Foot, Kidney, and Macrovascular complications.

Table 2: ICD-9 and ICD-10 codes used to define clinical complications in the study.

Results

The patient similarity model demonstrated competitive predictive performance. With an AUROC of 0.718, it outperformed Logistic Regression (0.695) and remained comparable to more complex models like Random Forest (0.764) and Support Vector Machines (0.766).

Beyond statistical performance, the model's primary value lies in its inherent transparency. A prototype web application was developed to visualize patient "timelines." In a case study involving a 40-year-old with poorly controlled diabetes, the prototype successfully identified similar patients whose outcomes could be used to encourage the patient to start insulin therapy—a narrative-based approach that is often more persuasive than a simple risk percentage.

Discussion and Conclusion

The patient similarity approach offers a significant shift from probability-based to narrative-based forecasting. By identifying historical "patients like you," the system provides clinicians with explainable evidence that aligns with medical intuition. Because the model's "explanation" is the data it retrieved (the 10 similar patients), there is no need for complex post-hoc explanation techniques.

While the model performed well on Singaporean polyclinic data, future work will focus on integrating such tools directly into EMR systems and assessing their impact on shared decision-making in live clinical environments. Ultimately, this approach can generate more informative CRPMs that better facilitate patient-provider communication and improve long-term health outcomes.

Summary updated

Abstract

Abstract Background: A Clinical Risk Prediction Model (CRPM) uses patient characteristics to estimate the probability about having or developing a particular disease and/or outcome. While CRPMs are gaining in popularity, they have yet to be adopted routinely in clinical practice. The lack of explainability and interpretability has limited its utility. Explainability is the extent of which a model’s prediction process can be described. Interpretability is the degree to which a user can understand the predictions made by a model.Methods: The study aimed to demonstrate utility of patient similarity analytics in developing an explainable and interpretable CPRM. Data was extracted from the electronic medical records of patients with type-2 diabetes mellitus, hypertension and dyslipidaemia in a Singapore public primary care clinic. We used various techniques, including patient similarity analytics, to develop various models on this real-world training dataset (n=7,041) and validated each of them on the same test dataset (n=3,018). The results were compared using logistic regression, random forest and support vector machine models from the same dataset. The CRPM was then implemented in a prototype system to demonstrate the identification, explainability and interpretability of similar patients and the prediction process.Results: The patient similarity model (AUROC=0.718) was comparable to the logistic regression (AUROC=0.695), random forest (AUROC=0.764) and support vector machine models (AUROC=0.766). We incorporated the patient similarity model in a prototype web application. A case study demonstrated how the application was provided both quantitative and qualitative information, in the form of patient narratives. This information was used to better inform and influence clinical decision-making, such as getting a patient to agree to start insulin therapy.Conclusions: A patient similarity approach is feasible to develop an explainable and interpretable CRPM. It is a general approach which can be used to develop locally relevant information, based on the database it searches. Ultimately, such an approach can generate a more informative CRPMs which can be deployed as part of clinical decision support tools to better facilitate shared decision-making in clinical practice.
BibTeX
@inproceedings{fang2021patient,
  author = {Hao Sen Andrew Fang and Ngiap Chuan Tan and Wei Ying Tan and Ronald Wihal Oei and Mong Li Lee and Wynne Hsu},
  title = {{Patient Similarity Analytics for Explainable Clinical Risk Prediction}},
  booktitle = {Research},
  year = {2021},
  doi = {10.21203/rs.3.rs-147408/v1},
}
Figure