Clinically Meaningful Comparisons Over Time: An Approach to Measuring Patient Similarity based on Subsequence Alignment

Dev Goyal, Zeeshan Syed, Jenna WiensarXiv 2018 · arXiv

Introduction

Comparing patient health trajectories over time is a cornerstone of managing chronic diseases like Alzheimer’s. However, traditional methods often fail to account for the inherent heterogeneity in disease progression. Researchers at the University of Michigan have proposed a novel approach based on subsequence alignment that significantly improves the accuracy of predicting disease transitions by allowing for flexible, data-driven comparisons of patient histories.

The Challenges of Longitudinal Analysis

Analyzing chronic disease data over time presents three primary hurdles that make standard chronological comparisons difficult:

  1. Pathophysiological Misalignment: Patients progress at different rates (e.g., "fast" vs. "slow" progressors), meaning two years post-diagnosis looks different for everyone.
  2. Variable Starting Points: Patients often enter clinical care at different stages of their disease progression.
  3. Irregular Sampling: Real-world clinical data is characterized by missed appointments and varying follow-up schedules, leading to gaps and inconsistent data lengths.
Visualization of patient visit patterns in the ADNI dataset.

Standard techniques like Global Dynamic Time Warping (DTW) attempt to align sequences from start to finish. In a clinical context, this is often inappropriate, as one patient's entire recorded history might only correspond to a small segment of another's longer trajectory.

Methodology: Subsequence Matching

The core of the proposed method is a flexible subsequence matching approach based on DTW. Instead of forcing a global match, this technique allows a shorter time series to align with any segment of a longer one, effectively "sliding" the trajectories to find the period of highest clinical similarity.

Overview of the subsequence extraction and prediction framework.

The model extracts multiple overlapping subsequences from each patient's history. Pairwise distances between these subsequences are calculated and used as features for an L2-regularized logistic regression classifier. To handle multi-modal data (MRI brain volume and PET glucose metabolism) where some scans may be missing, the authors employed matrix factorization to impute missing distance measures, ensuring no data was discarded due to incompleteness.

Experimental Evaluation

The model was evaluated using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), focusing on 511 patients with Mild Cognitive Impairment (MCI). The goal was to predict progression to "probable Alzheimer’s Disease" (AD) within a 36-month window.

The subsequence matching approach demonstrated superior performance compared to traditional baselines:

Method AUROC
Subsequence Matching 0.839
Global DTW 0.822
Snapshot (Most recent visit only) 0.812
Prefix Matching 0.813
Suffix Matching 0.820

The improvement over Global DTW and snapshot methods was statistically significant ($p < 0.01$). The model proved particularly effective at identifying the "pronounced decline" in brain volume that typically precedes an AD diagnosis, which single-visit snapshots often miss.

Conclusion

By prioritizing the rate of change and flexible alignment over absolute chronological values, this subsequence matching approach offers a more clinically meaningful way to compare patient trajectories. This data-driven alignment requires no specific "fiducial markers" (like a hospital admission date), making it a robust tool for monitoring chronic diseases with variable progression. Future work may extend this methodology to other conditions, such as diabetes or other neurodegenerative disorders, where clear temporal markers for alignment are missing.

Summary updated

Abstract

Longitudinal patient data has the potential to improve clinical risk stratification models for disease. However, chronic diseases that progress slowly over time are often heterogeneous in their clinical presentation. Patients may progress through disease stages at varying rates. This leads to pathophysiological misalignment over time, making it difficult to consistently compare patients in a clinically meaningful way. Furthermore, patients present clinically for the first time at different stages of disease. This eliminates the possibility of simply aligning patients based on their initial presentation. Finally, patient data may be sampled at different rates due to differences in schedules or missed visits. To address these challenges, we propose a robust measure of patient similarity based on subsequence alignment. Compared to global alignment techniques that do not account for pathophysiological misalignment, focusing on the most relevant subsequences allows for an accurate measure of similarity between patients. We demonstrate the utility of our approach in settings where longitudinal data, while useful, are limited and lack a clear temporal alignment for comparison. Applied to the task of stratifying patients for risk of progression to probable Alzheimer's Disease, our approach outperforms models that use only snapshot data (AUROC of 0.839 vs. 0.812) and models that use global alignment techniques (AUROC of 0.822). Our results support the hypothesis that patients' trajectories are useful for quantifying inter-patient similarities and that using subsequence matching and can help account for heterogeneity and misalignment in longitudinal data.
BibTeX
@misc{goyal2018clinically,
  author = {Dev Goyal and Zeeshan Syed and Jenna Wiens},
  title = {{Clinically Meaningful Comparisons Over Time: An Approach to Measuring Patient Similarity based on Subsequence Alignment}},
  year = {2018},
  eprint = {1803.00744},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1803.00744},
}
Figure