Imaging Pearls ❯ Deep Learning ❯ Electronic Health Records
|
-- OR -- |
|
- “The use of EHRs, along with the application of high-level computational techniques, have enabled researchers to query expansive (more than 100 million patients), multiorganizational clinical and administrative databases, such as Observational Health Data Sciences and Informatics, Patient-Centered Outcomes Research Network, and more recently, Epic Systems’ Cosmos. These databases, each representing a significant percentage of all patients, can enable large-scale representative retrospective observational studies to generate practice-based evidence regarding the delivery of safer, more equitable, and effective care. For example, researchers and policymakers can conduct postmarketing surveillance studies of adverse drug events and identify national patient safety trends.”
Maximizing the Ability of Health IT and AI to Improve Patient Safety.
Singh H, Sittig DF, Classen DC.
JAMA Intern Med. 2025 Jan 1;185(1):10-12. doi: 10.1001/jamainternmed.2024.4343. PMID: 39466271. - “Open-access health care data are scarce and lack geographicnand demographic diversity. Privacy constraints further limit data scope, reducing its effectiveness for underrepresented populations and specific conditions, which are inherently associated with a higher risk of reidentification. These limitations risk producing biased models that reinforce health care disparities. Synthetic data could help by generating additional data to support fairer predictions; for example, by selectively creating data for minority groups. However, creating realistic and useful medical data remains challenging, especially for rare conditions, dynamic changes, and outliers. In fields such as critical care, capturing intricate biological and clinical processes within highly detailed, multimodal data is difficult, potentially reducing the accuracy and reliability of AI generated insights. Synthetic data may also perpetuate or even amplify unresolved biases and spurious correlations from the original data, especially when produced at scale, which, without careful evaluation and mitigation, could exacerbate health care disparities rather than promote fairness and equity.”
Synthetic Data and Health Privacy
Gwénolé Abgrall, MD; Xavier Monnet, MD, PhD; Anmol Arora, MB, Bchir
JAMA Published online December 30, 2024 December 30, 2024. doi:10.1001/jama.2024.25821 - “Open-source models can offer greater transparency in managing privacy risks compared with closed platforms, as observed with Nabla’s recent shift from ChatGPT to open-source models, arguing for better control tailored to health care needs. Hosting large language models on private cloud infrastructure can further reduce risks, although it does not fully eliminate the chance of sensitive data leakage. Similarly,combining privacy-enhancing technologies, such as federated learning, synthetic data, and homomorphic encryption, can perhaps further strengthen data privacy while preserving utility.”
Synthetic Data and Health Privacy
Gwénolé Abgrall, MD; Xavier Monnet, MD, PhD; Anmol Arora, MB, Bchir
JAMA Published online December 30, 2024 December 30, 2024. doi:10.1001/jama.2024.25821