TL;DR
This paper introduces D3K, a novel patient similarity metric that combines electronic health record (EHR) data with clinical domain knowledge. By binning patient features according to medical guidelines and learning feature importance, D3K significantly outperforms baseline methods in identifying clinically similar patients, showing strong agreement with physician assessments.
What Problem Does This Paper Address?
Identifying patients with similar clinical profiles is crucial for personalized medicine, enabling targeted treatments and improved patient outcomes. However, existing patient similarity methods often struggle with the complexity and heterogeneity of Electronic Health Records (EHRs). Many approaches rely solely on data-driven insights, which may not align with clinical understanding, or are limited to datasets with low-level granularity (e.g., only ICD codes). This gap leaves a need for methods that can effectively leverage rich EHR data while incorporating established medical knowledge to identify truly clinically relevant patient similarities, especially for complex conditions like diabetes, hypertension, and dyslipidemia (DHL).
Methods
The core of this work is the D3K (Data-Driven and Domain Knowledge) patient similarity measure. D3K addresses the limitations of prior methods by:
- Comprehensive Feature Representation: It constructs patient feature vectors using a wide range of EHR data, including demographics, vital signs, laboratory test results, and prescribed medications (including dosages).
- Domain Knowledge-Based Binning: Instead of relying on traditional percentile-based binning, D3K discretizes continuous variables (like blood pressure, HbA1c, lipid levels) into bins that correspond to clinical guidelines and prognosis levels. This ensures that the bins reflect clinically meaningful distinctions. For example, blood pressure bins align with hypertension grading.
- Hybrid Feature Importance Learning: D3K learns the importance of features in two ways:
- Domain Knowledge: Physician-assigned labels are used to infer feature importance.
- Data-Driven: A generalized Mahalanobis distance is learned to maximize the distance between dissimilar patients and minimize it between similar ones.
- Cohort-Specific Similarity: The similarity computation is performed within specific sub-cohorts of patients defined by their comorbidities (e.g., diabetes only, hypertension and dyslipidemia). This allows for more localized and relevant similarity assessments.
- Similarity Calculation: The final similarity score between two patients is computed by taking the minimum of their binned scores for each variable, normalized by their average total scores.
Datasets
The study utilized a de-identified EHR dataset comprising 169,434 patients with diabetes, hypertension, or dyslipidemia (DHL) from seven primary care clinics in Singapore, collected between 2014 and 2015. This dataset was partitioned into seven distinct sub-cohorts based on the patients' specific DHL conditions and comorbidities.
Key Insights and Ideas
- Domain Knowledge is Crucial for Clinical Relevance: The most significant insight is that simply using data-driven binning (like percentiles) is insufficient. Aligning feature discretization with established clinical guidelines leads to patient similarity measures that are more meaningful to clinicians.
- Comorbidities Matter: Retrieving similar patients within specific sub-cohorts (defined by comorbidities) yields better results than searching across the entire patient population. This highlights the importance of considering the complex interplay of diseases.
- Comprehensive Data is Key: Including a wide array of data types, from demographics to medication dosages, provides a richer picture of patient health and leads to more accurate similarity assessments.
- Physician Agreement Validates the Approach: The substantial agreement between D3K's results and physician judgments is a strong indicator that the method captures clinically relevant similarities.
Key Results
- Superior Performance: D3K achieved the highest nDCG@10 (a measure of retrieval effectiveness) across all seven sub-cohorts, significantly outperforming Euclidean distance and Locally Supervised Metric Learning (LSML). For example, in the CDHL cohort, D3K achieved an nDCG of 0.831, compared to 0.809 for Euclidean distance and 0.781 for LSML.
- Domain-Based Binning Wins: The domain knowledge-based binning strategy consistently outperformed percentile-based binning in most cohorts, with statistically significant improvements in several cases.
- Physician Agreement: Cohen's kappa showed substantial agreement between D3K and physicians (κ = 0.715 and κ = 0.863), and Fleiss' kappa indicated significant agreement between both physicians and D3K (κ = 0.746). This suggests D3K can be a valuable tool for shared decision-making.
- Sub-cohort Retrieval Advantage: Retrieving similar patients from within their specific sub-cohorts generally resulted in higher nDCG@10 compared to retrieving from the entire study cohort.
Limitations and Future Work
- Dataset Specificity: The study was conducted on a dataset from Singapore. While the approach is generalizable, its performance might vary on datasets from different healthcare systems or patient populations.
- Temporal Trajectory: The current D3K model primarily uses data from a single "base visit." The authors suggest future work should consider the temporal trajectory of patients over multiple visits to capture disease progression and treatment responses more dynamically.
- Data Augmentation: Expanding the dataset to include patients from multiple clinic sites would enhance the robustness and generalizability of the model.
- Downstream Applications: The authors envision D3K as a personalized clinical decision tool and suggest exploring its application in other downstream tasks that benefit from patient similarity.
Summary updated