1143 Innovations in Time Series Data Integration in a Polysomnographic Foundational Model Informing Risk Groups

Le résumé fourni par la source

Abstract Introduction Despite the abundance of polysomnography (PSG) data, the limited summary metrics used in existing approaches may not provide the most informative insights for clinical decision-making. We hypothesized that a new data-driven clustering method using the entire multimodal raw PSG data could enable a precise risk stratification approach. We leveraged a new clinical data set to facilitate this data-driven approach and create a novel Foundation Model. Methods We utilized 10,000 PSGs conducted at the Cleveland (1/2012-12/2022) and custom artificial intelligence techniques that incorporate time-series data, to develop a Foundation Model from raw PSG data. We optimized this new model to classify sleep stages, respiratory events, and oxygen desaturations. Resulting embeddings were used to cluster patients into distinct risk groups with a k-means algorithm. Baseline characteristics were compared between risk groups using chi-square tests for categorical variables and Welch’s ANOVA for continuous variables. Means and standard deviations are reported. Results Optimal stratification of embeddings was achieved with five clusters. Resulting risk groups (RG) had a graded increase in age, male predominance, cardiovascular risk factors, and sleep-disordered breathing(SDB) severity from RG1(n=3,357) to RG5(n=363). Males were more prevalent in RG4(n=1,144; 60.8%) and RG5(66.4%) and least in RG2(n=1,877; 37.6%). Body mass index was lowest in RG2(32.7±9.1kg/m2) and highest in RG5(35.4±10.4kg/m2). RG4 and RG5 were the oldest(58.6±16.1years) and RG2 the youngest(44.0±15.0years). RG1 had intermediate SDB with apnea hypopnea index (AHI:12.4±12.4), total sleep time (TST:328±61.1min), and risk (hypertension:59.8%,diabetes II:32.7%), and lowest cognitive impairment(14.1%). RG2 had the mildest SDB (AHI:5.4±6.4), longest TST (342±71.6min), and lowest cardiovascular risk (hypertension:47.5%,diabetes II:24.7%). RG3(n=2,867) had intermediate SDB and risk. RG4 had more abnormal PSG measures (AHI:22.7±13.2, TST:201±73.9min), more risk (hypertension:75.6%,diabetes II:65.2%), and highest major adverse cardiovascular events (43.6%) but lowest migraine (11.5%). RG5 had the most severe SDB (AHI:37.3±39.0), lowest TST (98.4±89.0min), and high risk (hypertension:77.4%,diabetes II:44.6%). All p-values were < 0.001. Conclusion We created a Polysomnographic Foundational Model to stratify patients into risk groups characterized by different comorbidities not completely explained by traditional measures. RG4 and RG5 exhibited more severe SDB and unique clinical characteristics that prompt future investigation and warrant more attention from healthcare providers. Support (if any) IBM Discovery Accelerator, AIM Award, NIH 1R21HL170206-01

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.