Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Computing patient similarity based on unstructured clinical notes

Petr Zelina, Marko Řeháček, Jana Halámková, Lucia Bohovicová, Martin Rusinko, Vít Nováček

TL;DR

The paper addresses the challenge of computing patient similarity from unstructured EHR notes by representing each patient as a matrix of per-note embeddings and evaluating category-specific similarity. It introduces a modular pipeline with segmentation/filtering, multiple embedding methods (LSA, Doc2Vec, sentence-BERT), and three matrix similarity measures (RV coefficient, MaxMax, Edit-Distance), validated against expert annotations. A 50-pivot validation study across 10 clinical categories shows that certain configurations, particularly RV2 with LSA or a combined embedding approach, provide the most reliable similarity for treatment-related categories, though several facets (e.g., social history, allergies) remain challenging. The work demonstrates practical utility for downstream tasks such as personalized therapy recommendations and toxicity warnings, while highlighting the importance of category-focused filtering and robust validation design; the codebase is publicly available for reproduction.

Abstract

Clinical notes hold rich yet unstructured details about diagnoses, treatments, and outcomes that are vital to precision medicine but hard to exploit at scale. We introduce a method that represents each patient as a matrix built from aggregated embeddings of all their notes, enabling robust patient similarity computation based on their latent low-rank representations. Using clinical notes of 4,267 Czech breast-cancer patients and expert similarity labels from Masaryk Memorial Cancer Institute, we evaluate several matrix-based similarity measures and analyze their strengths and limitations across different similarity facets, such as clinical history, treatment, and adverse events. The results demonstrate the usefulness of the presented method for downstream tasks, such as personalized therapy recommendations or toxicity warnings.

Computing patient similarity based on unstructured clinical notes

TL;DR

The paper addresses the challenge of computing patient similarity from unstructured EHR notes by representing each patient as a matrix of per-note embeddings and evaluating category-specific similarity. It introduces a modular pipeline with segmentation/filtering, multiple embedding methods (LSA, Doc2Vec, sentence-BERT), and three matrix similarity measures (RV coefficient, MaxMax, Edit-Distance), validated against expert annotations. A 50-pivot validation study across 10 clinical categories shows that certain configurations, particularly RV2 with LSA or a combined embedding approach, provide the most reliable similarity for treatment-related categories, though several facets (e.g., social history, allergies) remain challenging. The work demonstrates practical utility for downstream tasks such as personalized therapy recommendations and toxicity warnings, while highlighting the importance of category-focused filtering and robust validation design; the codebase is publicly available for reproduction.

Abstract

Clinical notes hold rich yet unstructured details about diagnoses, treatments, and outcomes that are vital to precision medicine but hard to exploit at scale. We introduce a method that represents each patient as a matrix built from aggregated embeddings of all their notes, enabling robust patient similarity computation based on their latent low-rank representations. Using clinical notes of 4,267 Czech breast-cancer patients and expert similarity labels from Masaryk Memorial Cancer Institute, we evaluate several matrix-based similarity measures and analyze their strengths and limitations across different similarity facets, such as clinical history, treatment, and adverse events. The results demonstrate the usefulness of the presented method for downstream tasks, such as personalized therapy recommendations or toxicity warnings.
Paper Structure (21 sections, 4 figures, 3 tables)

This paper contains 21 sections, 4 figures, 3 tables.

Table of Contents

  1. Similarity computation pipeline
  2. Pipeline overview
  3. Clinical note segmentation and filtering
  4. Clinical note vectorization
  5. Patient matrix similarity
  6. Clinical notes dataset
  7. Experiments
  8. Explored parameter space
  9. Performance
  10. Validation study
  11. Similarity categories
  12. Study design
  13. Results
  14. Evaluation methodology
  15. Result data
  16. ...and 6 more sections

Figures (4)

  • Figure 1: Overview of our methodology. Adapted from Zelina2023thesis.
  • Figure 2: Example of patient matrices of a patient with 62 clinical notes (rows). The embedding dimension is 50 (columns). Zelina2023thesis
  • Figure 3: Comparison of matrix similarity methods of synthetic patient matrices. The underlying "patient" matrices were generated so that the similarity matrices look in a certain way. Matrices: Similarity matrix of two synthetic patients. Red lines: the best path found by the eds algorithm. Table under matrices: Similarity scores for given synthetic patient pair. Zelina2023thesis
  • Figure 4: Detailed performance analysis of the Rrv2-Vlsa050 combination (best performing). The boxplot shows a comparison of the inter-annotator agreement (blue), the performance of the model without filtering (orange, ✗), and the performance of the model with filtering (green, ✓) for different categories.