#5490 GENERATIVE ARTIFICIAL INTELLIGENCE FOR CREATION OF SYNTHETIC HYPERTENSION TRIAL DATA

Abstract Background and Aims Synthetic data can be an effective supplement or alternative to real for the training of machine learning models. may also used evaluate new tools, develop educational curricula, remove undesirable biases in datasets. We aim four synthetic generation methods applied hypertension randomized clinical trial data. Method The Systolic Blood Pressure Intervention Trial (SPRINT) showed that intensive BP control SBP <120 mm Hg results significant cardiovascular benefits high-risk patients with compared routine <140 Hg. Data Vault (SDV) is a Generation ecosystem libraries allows users easily generate has same format statistical properties as original dataset. SDV supports multiple types data, including date-times, discrete-ordinal, categorical, numerical. SPRINT was pre-processed create single table 140,000 patient visits baseline variables (age, sex, race, aspirin use, estimated Glomerular Filtration Rate (eGFR)) visit level (systolic diastolic blood pressure, heart rate total number antihypertensive medications at end visit). Using library python, we generative models 1. Gaussian copula model, 2. Conditional Tabular Generative adversarial network (CTGAN), 3. CopulaGan 4. Variational Auto-encode (TVAE). evaluated using SDMetrics which includes shapes columns (marginal distributions), pairwise trends between (correlations), reproduce mathematical from your row synthesis. Finally, overall quality score represents amalgamation marginal distribution correlations computed, where 0 indicates lowest 1 highest. Results Two hundred thousand were created each method. scores order 90.67% copula, 86.77% TVAE, 81.03% CTGAN’, 79.7% CopulaGAN. column shape highest Copula (94.54%), followed by TVAE (88.44%), CTGAN (82.35%), GAN (80.27%). pair trend corresponds (86.8%), TAVE (85.1%), (79.72%), (79.12%). Conclusion scoring based on distribution, correlations, score. feasible collection future AI