Medical records condensation: a roadmap towards healthcare data democratisation

Anshul Thakur, Yujiang Wang, Mingzhi Dong, Pingchuan Ma, Stavros Petridis, Li Shang, Tingting Zhu, David A. CliftonResearch Square 2024 · DOI

Abstract

The prevalence of artificial intelligence (AI) has envisioned an era of healthcare democratisation that promises every stakeholder a new and better way of life. However, the advancement of clinical AI research is significantly hurdled by the dearth of data democratisation in healthcare. To truly democratise data for AI studies, challenges are two-fold:

  1. the sensitive information in clinical data should be anonymised appropriately, and 2. AI-oriented clinical knowledge should flow freely across organisations. This paper considers a recent deep-learning advent, dataset condensation (DC), as a stone that kills two birds in democratising healthcare data. The condensed data after DC, which can be viewed as statistical metadata, abstracts original clinical records and irreversibly conceals sensitive information at individual levels; nevertheless, it still preserves adequate knowledge for learning deep neural networks (DNNs). More favourably, the compressed volumes and the accelerated model learnings of condensed data portray a more efficient clinical knowledge sharing and flowing system, as necessitated by data democratisation. We underline DC's prospects for democratising clinical data, specifically electrical healthcare records (EHRs), for AI research through experimental results and analysis across three healthcare datasets of varying data types.

The Challenge of Healthcare Data Sharing

The "Achilles heel" of clinical AI research is the lack of accessible data. While AI promises to revolutionize healthcare, two major hurdles persist:

  1. Privacy Risks: Clinical data is highly sensitive. Traditional anonymization (redacting names/IDs) is often insufficient, as sophisticated attacks can still re-identify patients from patterns in their medical records.
  2. Resource Intensity: Electronic Health Records (EHRs) are massive. Maintaining, transporting, and training models on these datasets is economically and computationally expensive, creating a high barrier to entry for researchers without massive GPU clusters.

The authors propose that we stop trying to share raw, redacted records and instead share "condensed" versions that function as statistical metadata.

Methodology: Dataset Condensation (DC)

The core technique is Dataset Condensation (DC) with distribution matching. Unlike standard compression, which tries to save space by removing redundancy, DC aims to synthesize a brand-new, much smaller dataset that "teaches" a neural network the same knowledge as the original.

  • Distribution Matching: The algorithm seeks to minimize the discrepancy between the original data and the condensed data in a high-dimensional embedding space.
  • Randomized Encrypters: To ensure privacy, the authors use a cohort of randomly initialized Deep Neural Networks (DNNs) to project the data. Because these networks are never fully "trained" on the raw data and their parameters are discarded after condensation, they act as innate encrypters.
  • Many-to-One Mapping: Each condensed sample is a synthesis of information from many original patients. This creates a "bona fide" de-identification because there is no longer a one-to-one correspondence between a synthetic record and a real person.

Experimental Results and Analysis

The authors validated their approach across three diverse healthcare datasets: PhysioNet-2012 (ICU mortality prediction), MIMIC-III (in-hospital death prediction), and Coswara (COVID-19 respiratory sound diagnosis).

Preserved Utility and Model Generalizability

Across all 11 different DNN architectures tested, the condensed data consistently allowed models to achieve high performance. This proves that the "knowledge" captured is generalizable and not overfit to one specific model type.

Test AUC across 11 DNN architectures for original and condensed PhysioNet-2012 datasets.

Massive Compression and Efficiency

For PhysioNet-2012, the authors condensed 88.13 MB of data (5,120 samples) into just 0.17 MB (20 samples)—a 99.8% reduction in size—while maintaining a Test AUC of approximately 0.80 (compared to the original 0.858). Furthermore, models trained on condensed data reach peak performance significantly faster. In the MIMIC-III experiments, training steps were reduced from 1,860 to just 81—a 95.6% speedup.

Trade-off between predictive performance (Test AUC) and data storage size (Megabytes) for different condensation levels.

Conclusion

While DC isn't a "silver bullet" for every analytical task, it provides a powerful new tool for researchers to share high-utility, low-risk datasets. By reducing data volumes by over 99% and accelerating model training by up to 95%, DC offers a practical roadmap toward the democratisation of clinical AI research.

Summary updated

Abstract

<title>Abstract</title> The prevalence of artificial intelligence (AI) has envisioned an era of healthcare democratisation that promises every stakeholder a new and better way of life. However, the advancement of clinical AI research is significantly hurdled by the dearth of data democratisation in healthcare. To truly democratise data for AI studies, challenges are two-fold: 1. the sensitive information in clinical data should be anonymised appropriately, and 2. AI-oriented clinical knowledge should flow freely across organisations. This paper considers a recent deep-learning advent, dataset condensation (DC), as a stone that kills two birds in democratising healthcare data. The condensed data after DC, which can be viewed as statistical metadata, abstracts original clinical records and irreversibly conceals sensitive information at individual levels; nevertheless, it still preserves adequate knowledge for learning deep neural networks (DNNs). More favourably, the compressed volumes and the accelerated model learnings of condensed data portray a more efficient clinical knowledge sharing and flowing system, as necessitated by data democratisation. We underline DC's prospects for democratising clinical data, specifically electrical healthcare records (EHRs), for AI research through experimental results and analysis across three healthcare datasets of varying data types.
BibTeX
@inproceedings{thakur2024medical,
  author = {Anshul Thakur and Yujiang Wang and Mingzhi Dong and Pingchuan Ma and Stavros Petridis and Li Shang and Tingting Zhu and David A. Clifton},
  title = {{Medical records condensation: a roadmap towards healthcare data democratisation}},
  booktitle = {Research Square},
  year = {2024},
  doi = {10.21203/rs.3.rs-3781993/v1},
}
Figure