Balancing Privacy and Utility in Child and Adolescent Mental Health Services Research: Retrospective Cohort Study on Synthetic Data Generation.
Electronic health records are essential for advancing research aimed at improving clinical outcomes. However, stringent data protection and privacy concerns severely limit the accessibility and use of real clinical data, particularly within Child and Adolescent Mental Health Services (CAMHS) involving vulnerable young individuals. This challenge can be effectively addressed through synthetic data generation, which safeguards individual privacy while facilitating comprehensive analyses of clinical information.
This study aims to investigate whether hierarchical synthetic data generators (SDGs) can effectively replicate the statistical properties, preserve the utility, and maintain the privacy of real CAMHS clinical data, thereby enabling data sharing and broader access to research-ready datasets.
This retrospective cohort study used electronic medical record data from 6924 distinct patients from CAMHS in Stavanger, Norway, comprising 7730 referral periods and 58,524 episodes of care. An 80%-20% split was used for training and testing. A hierarchical synthetic data generation model was trained to generate synthetic referral periods and associated episodes of care. Data quality was evaluated using SDMetrics for distribution (Kolmogorov-Smirnov Complement [KSC]/Total Variation Complement [TVC]), correlation (CorrelationSimilarity [CS]), and cardinality (CardinalityShapeSimilarity [CSS]) similarity. Privacy was evaluated using the Anonymeter library to simulate singling out, linkability, and inference reidentification attacks. Utility was assessed using the train synthetic test real (TSTR) pattern, comparing the predictive performance using precision-recall area under the curve [PRAUC] of models trained on synthetic vs real data for classifying the intensity of care.
The hierarchical SDG created highly reproducible synthetic CAMHS data. The average statistical similarity scores were high across all metrics: KSC/TVC at 0.92, CS at 0.77 (intertable CS at 0.75), and CSS at 0.92. The synthetic data also demonstrated a low risk under simulated privacy attacks on a control dataset (n=1546): the average success rate was 6/1546 (0.39%) for singling out and 77/1546 (5%) for multivariate attacks. The average linkability risk was 54/1546 (0.5%), and the highest inference risk for a sensitive variable was 2/1546 (0.12%). The classification model trained on synthetic data (TSTR) produced comparable predictive performance (PRAUC=0.40) to the model trained on real data (PRAUC=0.43) for classifying the intensity of care (low vs medium or higher). Shapley additive explanations analysis confirmed that the synthetic model's explanations aligned with real-world insights, validating its ability to capture fundamental predictive patterns.
Synthetic data can be used to build trust and promote collaboration among CAMHS researchers by offering access to extensive, representative samples with a low risk of patient identification. This approach expands the breadth of research while safeguarding patient privacy. Effective implementation of synthetic data generation depends on the model's ability to accurately identify and replicate the complex, sequential patterns present in real data.
This study aims to investigate whether hierarchical synthetic data generators (SDGs) can effectively replicate the statistical properties, preserve the utility, and maintain the privacy of real CAMHS clinical data, thereby enabling data sharing and broader access to research-ready datasets.
This retrospective cohort study used electronic medical record data from 6924 distinct patients from CAMHS in Stavanger, Norway, comprising 7730 referral periods and 58,524 episodes of care. An 80%-20% split was used for training and testing. A hierarchical synthetic data generation model was trained to generate synthetic referral periods and associated episodes of care. Data quality was evaluated using SDMetrics for distribution (Kolmogorov-Smirnov Complement [KSC]/Total Variation Complement [TVC]), correlation (CorrelationSimilarity [CS]), and cardinality (CardinalityShapeSimilarity [CSS]) similarity. Privacy was evaluated using the Anonymeter library to simulate singling out, linkability, and inference reidentification attacks. Utility was assessed using the train synthetic test real (TSTR) pattern, comparing the predictive performance using precision-recall area under the curve [PRAUC] of models trained on synthetic vs real data for classifying the intensity of care.
The hierarchical SDG created highly reproducible synthetic CAMHS data. The average statistical similarity scores were high across all metrics: KSC/TVC at 0.92, CS at 0.77 (intertable CS at 0.75), and CSS at 0.92. The synthetic data also demonstrated a low risk under simulated privacy attacks on a control dataset (n=1546): the average success rate was 6/1546 (0.39%) for singling out and 77/1546 (5%) for multivariate attacks. The average linkability risk was 54/1546 (0.5%), and the highest inference risk for a sensitive variable was 2/1546 (0.12%). The classification model trained on synthetic data (TSTR) produced comparable predictive performance (PRAUC=0.40) to the model trained on real data (PRAUC=0.43) for classifying the intensity of care (low vs medium or higher). Shapley additive explanations analysis confirmed that the synthetic model's explanations aligned with real-world insights, validating its ability to capture fundamental predictive patterns.
Synthetic data can be used to build trust and promote collaboration among CAMHS researchers by offering access to extensive, representative samples with a low risk of patient identification. This approach expands the breadth of research while safeguarding patient privacy. Effective implementation of synthetic data generation depends on the model's ability to accurately identify and replicate the complex, sequential patterns present in real data.
Authors
Haizoune Haizoune, Leventhal Leventhal, Pant Pant, Nytrø Nytrø, Koochakpour Koochakpour, Koposov Koposov, Øhlckers Øhlckers, Skokauskas Skokauskas
View on Pubmed