Computational Phenotyping Research Articles

1318-P: Predicting Future Development of Type 2 Diabetes Using the EHR-Driven Dysglycemia Risk Score (D-RISK)

Introduction and Objective: The Dysglycemia Risk Score (DRISK) is a validated, EHR-based risk score to identify patients at high risk for undiagnosed dysglycemia (prediabetes and T2D). DRISK performs better than screening guidelines to detect prevalent, undiagnosed dysglycemia; however, its ability to predict future development of T2D is unknown. Methods: Participants in a T2D screening study, excluding those with T2D on baseline screening, were analyzed. DRISK score (range 0-28; components: age, race, BMI, hypertension, random blood glucose (RBG)) was calculated from EHR data. Incident T2D was defined using a validated, EHR computable phenotype (2 of 3: HbA1c > 6.5%, T2D diagnosis codes; antihyperglycemic medications) calculated after the initial screening visit until the patients’ last EHR encounter. We used the baseline DRISK score to predict incident T2D using a Cox proportional hazard model with censoring for death or loss to followup. We adjusted for baseline HbA1c to address potential confounding during followup. Results: A total of 502 individuals without T2D at baseline (mean age 48y; BMI 30 kg/m2; 70% female; 69% Hispanic, 22% Black; 35% hypertension; 31% prediabetes (HbA1c≥5.7%); mean DRISK score 9.5) were analyzed. Over a mean followup of 5.8 years, 12% progressed to T2D. Those having higher BMI, hypertension, prediabetes, and higher HbA1c, RBG and DRISK scores at baseline were more likely to develop T2D (p<0.05 for all). Patients with DRISK scores ≥10 were more likely to progress to T2D than those with DRISK scores <10 (HR=1.9, p<0.015). ADA and USPSTF screening guidelines (HR 1.4 and 1.3 respectively; p>0.05 for both) did not predict future development of T2D in this cohort. Conclusion: DRISK can identify individuals at high risk of progression to T2D and has greater prognostic value than commonly used screening guidelines. By utilizing routinely available, structured EHR data, DRISK may help identify patients at high risk of progressing to T2D within health systems. Disclosure A. Mamun: None. M. McGuire: None. V. Merrill: None. S. Zhang: None. N.O. Santini: None. B. Moran: None. L. Meneghini: Employee; Sanofi. Stock/Shareholder; Sanofi. I. Lingvay: Consultant; Abbvie, Altimmune, Amgen, Alveus Tx, Antag Tx, Astra Zeneca, Bayer, Betagenon AB, Bioio Inc., Biomea, Boehringer-Ingelheim, Carmot, Cytoki Pharma, Eli Lilly, Intercept, Janssen/J&J, Juvena, Keros Ther, Novo Nordisk, Pharmaventures, Pfizer, Regeneron, Roche, Sanofi, Shionogi, Source Bio, Structure Therapeutics, TARGET RWE, TERNS Pharma, The Comm Group, WebMD, and Zealand Pharma. Research Support; Novo Nordisk, Sanofi, Boehringer-Ingelheim. E. Halm: None. M.E. Bowen: Research Support; Boehringer-Ingelheim. Funding National Institute of Diabetes Digestive and Kidney Diseases at NIH (K23DK104065)

399-P: Towards a Digital Twin Model for Predicting Renal Function Decline in Type 2 Diabetes Using Longitudinal Electronic Health Records

Introduction and Objective: T2D is a public health crisis and a leading cause of end-stage kidney disease (ESKD). This study aims to develop T2DTwin, a digital twin model to simulate T2D progression and predict complications, with a focus on modeling diabetic kidney disease progression using longitudinal electronic health records (EHRs). Methods: We identified adult T2D patients from the 2012-2021 OneFlorida+ longitudinal EHR data via a validated computable phenotyping algorithm, including those with at least 2 eGFR values during follow-up. Our T2DTwin model utilizes an Operator-Regression-based machine learning (ML) approach to learn complex dynamical systems from data, integrating time-varying and time-invariant information (e.g., sociodemographic, vitals, lab values, comorbidities, and medications). We implemented neural operators to model GFR to (1) forecast renal function trajectories (e.g., eGFR decline), and (2) assess the impact of modifiable predictors, like SGLT2 inhibitors. Results: Among 251,305 T2D patients (mean age 58.3 years), baseline eGFR ranged from 30.0 to 126.8 mL/min/1.73 m2, with a mean (SD) of 72.5 (23.6). T2DTwin showed strong predictive utility for eGFR changes over one year, achieving SMAPE of 10.6% (i.e., 10.6% off from the actual eGFR). We predicted risks of significant renal function decline (e.g., 30% eGFR reduction within 1, 3, and 5 years) and CKD stage changes (1-5), with F1 score, precision, and sensitivity of 0.78. SHAP analysis identified triglycerides and low density lipoprotein as important modifiable factors. Conclusion: T2DTwin leverages Operator-Regression to model complex dynamic systems like human body, offering scalability and adaptability over traditional ML methods. Its accurate predictions of renal function decline in T2D patients using longitudinal EHR have significant potential to enhance diabetic kidney disease management. Disclosure H. Chen: None. Y. Huang: None. L. Sun: None. A. Chen: None. J. Guo: None. J. Bian: None.

Comparative effectiveness of epilepsy surgery versus additional anti-seizure medications for Lennox-Gastaut syndrome: study protocol for a multicenter, mixed-methods study.

Lennox-Gastaut Syndrome (LGS) is a severe developmental epileptic encephalopathy without a known cure. Management of symptoms requires substantial care. Treatment options include anti-seizure medications, dietary therapy, and epilepsy surgery. Two main treatment pathways for patients with LGS with drug resistant epilepsy are additional anti-seizure medications or epilepsy surgery, which have been reported to be effective in reduction of seizure burden and improving quality of life. No studies have directly compared the outcomes of using epilepsy surgery versus using additional anti-seizure medications for the treatment of LGS. This study is a multicenter, mixed-methods comparative effectiveness study of LGS patients who have undergone epilepsy surgery or have received an LGS-approved medication for treatment resistant seizures. Aim 1 will analyze the effect of surgical therapies and additional medication on two clinical outcomes: (1a) seizure-related healthcare utilization, and (1b) expressive communication, behavior, and parent-reported quality of life. Based on electronic health record review and coding validation as part of Aim 1a, we will develop computable phenotypes for LGS. The phenotypes will inform the analyses in Aim 1a and Aim 2. Aim 2 will describe the real-world utilization of these treatments across multiple healthcare institutions in the United States. Data will be collected from electronic health records, data marts in the National Patient-Centered Clinical Research Network (PCORnet®) format, caregiver surveys, and focus groups. This study of LGS will provide currently unavailable evidence concerning the real-world comparative effectiveness of epilepsy surgeries and additional anti-seizure medications. The outcomes are those that families identify as important: emergency medical care for seizures and patients' functional outcomes. The results of this study may help guide decisions regarding the treatment of LGS and development of computable phenotypes for this rare disease. This study using PCORnet® data will also lay the groundwork for future large-scale studies on LGS and other rare epilepsies. ClinicalTrials.gov, identifier NCT05374824.

2117-LB: Comparing Computable Phenotypes for Identification of Diabetes in an Integrated Health Care System

Introduction and Objective: Conducting rapid diabetes surveillance across health systems using electronic health records (EHR) can potentially replace costly manual chart reviews. However, such surveillance requires a parsimonious case-finding algorithm with high predictive value. Our objective was to determine an algorithm with good performance using EHR. Methods: We conducted chart reviews of 627 Kaiser Permanente Southern California members age 18-44 years, randomly selected from a subgroup of 55,089 patients with presumed diabetes identified by any diabetes-related diagnosis code, medication, or abnormal laboratory result using EHR in 2018. With chart review as the gold standard, we compared the performance of several algorithms when defining diabetes computable phenotypes, indicated by sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). Performance was examined overall and by diabetes type, sex and race-ethnicity. Results: Identifying diabetes with diagnosis codes alone produced the best performance with sensitivity, specificity, PPV, and NPV of 94% [95% Confidence Interval: 91%-95%]; 87% (78%-93%); 98% (97%-99%) and 69% (42%-76%), respectively. When requiring a diagnosis code plus an abnormal lab or a diagnosis code plus diabetes medication, sensitivity dropped to 39% (35%-44%) and 44% (40%-49%), respectively, but specificity increased to 100%. When using a less restrictive algorithm such as requiring an abnormal lab or antidiabetic medication with a diagnosis code, sensitivity increased to 98% (97%-99%) and 97% (95%-98%), respectively, while the specificity dropped to under 50% (range: 0%-48%). Using only a diagnosis code resulted in overall good performance in identifying both T1D and T2D but the sensitivity of identified T1D was lowest among Black individuals [63% (25%-92%)]. We observed no difference in performance by sex. Conclusion: A simple algorithm using a diagnosis code only from the EHR demonstrates good performance in identifying both T1D and T2D diabetes. Disclosure H. Zhou: None. M.M. Zhou: None. D. McCarthy: None. T. Harrison: None. M.T. Mefford: Research Support; Merck & Co., Inc. J. Chang: None. N. Mensah: None. J.M. Chang: None. K. Reynolds: Research Support; Merck & Co., Inc.

Development and validation of a computable phenotype for adolescent idiopathic scoliosis

Abstract IntroductionThere remains a lack of understanding of the etiology and treatment effectiveness for Adolescent idiopathic scoliosis (AIS). The objective of this study was to develop and validate a computable phenotype for patients with AIS to facilitate rapid learning through large‐scale observational research using real‐world data.Study DesignFour computable phenotype (CP) algorithms were developed and tested. The algorithms were executed against the Shriners Children's (SC) Research Data Warehouse using the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) from January 1, 2016 to December 31, 2022. CPs composed of diagnosis and imaging procedure utilization codes were evaluated iteratively against a prospective registry of scoliosis patients. The highest‐performing phenotype was then evaluated through manual chart review for validation. Demographic characteristics of the patients meeting the phenotype definition were assessed.ResultsThe four alternative CPs ranged from 24 103 to 15 292 unique patients. The CP that balanced sensitivity (92.7%) and specificity (81.8%) when evaluated against a prospective registry of scoliosis patients was chosen as the final AIS CP. Among 50 patients with phenotype‐confirmed AIS, 36 (72%) had chart‐validated AIS, and 14 (28%) were identified as false positives. Of the 14 false positives, 6 cases had a diagnosis of spinal asymmetry. Among the patients meeting the phenotype definition, the average age of patients with AIS treated at SC is 13.6 years (SD = 1.64) and patients are primarily female (73.7%) and white (56.2%).ConclusionThe CP had good performance in identifying pediatric patients with AIS. Future refinements to the algorithm should include the use of x‐ray parameters or the application of natural language processing to unstructured EHR data to better distinguish AIS cases from other spinal diagnoses. This CP is a fundamental step to facilitate a learning health system environment that can rapidly develop evidence to improve pediatric patient outcomes.

Identification of Transgender and Gender-Diverse Individuals in the All of Us Research Program, 2017-2022.

Objectives. To develop computable phenotype algorithms to identify a transgender and gender-diverse (TGD) cohort by using diverse data sources in All of Us, a national community-engaged program to facilitate health equity in the United States by partnering with 1 million participants. Methods. We identified TGD individuals in All of Us by applying inclusion criteria based on conditions, laboratory measurements, or medications related to being TGD in electronic health record data or confirmed survey responses, using participant data collected between May 31, 2017, and July 1, 2022. Results. Of 413 457 participants, we identified 4781 (1.2%) as TGD. Participants aged 18 to 29 years (26.1% vs 8.2%), who were bisexual (20.7% vs 3.5%), with annual income of less than $25 000 (35.9% vs 24.7%), and with housing security concerns (31.9% vs 16.0%) accounted for a larger proportion of TGD individuals than non-TGD individuals. Conclusions. Combining survey and electronic health record data enables the identification of TGD individuals who have been missed by previous studies that used survey data alone in All of Us to explore health disparities in TGD people. (Am J Public Health. Published online ahead of print June 5, 2025:e1-e10. https://doi.org/10.2105/AJPH.2025.308129).

Utilization of anti-CD20 antibodies for treatment of childhood nephrotic syndrome, 2010 to 2022.

A growing body of evidence supports the efficacy of the type I anti-CD20 monoclonal antibody, rituximab, in the management of children with frequently relapsing or steroid-dependent nephrotic syndrome. We examined temporal trends and described current patterns in the use of anti-CD20 antibodies and other corticosteroid-sparing drug therapies in a large multi-institutional population of children with nephrotic syndrome. Data came from PEDSnet, a clinical research network that aggregates electronic health record data at several children's healthcare organizations in the United States. Patients with at least one inpatient, emergency, or outpatient physician encounter between January 2010 and November 2022 who met our published computable phenotype algorithm for nephrotic conditions were included. Children with systemic lupus erythematosus or congenital/genetic nephrotic diagnoses were excluded. Treatments were measured from nephrotic syndrome diagnosis to kidney transplant or most recent encounter. Among 6,892,137 patients across 6 centers, 2962 met criteria for nephrotic conditions (0.4 per 1000 patients). 852 (28.8%) had at least one native kidney biopsy. Nearly half of the population was exposed to at least one steroid-sparing agent, most of whom had exposure to multiple agents. 524 (17.7%) patients were exposed to rituximab, and utilization of rituximab increased over the 12-year study period. Similar trends were observed for mycophenolate and tacrolimus. Concurrently, use of cyclosporine and cyclophosphamide decreased. Use of rituximab to manage nephrotic syndrome has steadily increased, and tacrolimus, mycophenolate, and rituximab are currently the most commonly used steroid-sparing agents for childhood nephrotic syndrome.

Computational Phenotyping of Effort-Based Decision Making in Unmedicated Adults With Remitted Depression.

Computational Phenotyping of Effort-Based Decision Making in Unmedicated Adults With Remitted Depression.

Variants in BSN, encoding the presynaptic protein Bassoon, result in a distinct neurodevelopmental disorder with a broad phenotypic range.

Variants in BSN, encoding the presynaptic protein Bassoon, result in a distinct neurodevelopmental disorder with a broad phenotypic range.

Development and Evaluation of a Computable Phenotype for Normal Tension Glaucoma

Development and Evaluation of a Computable Phenotype for Normal Tension Glaucoma

Identifying patients with neurofibromatosis type 1 related optic pathway glioma using the OMOP CDM.

Identifying patients with neurofibromatosis type 1 related optic pathway glioma using the OMOP CDM.

Computable phenotypes to identify respiratory viral infections in the All of Us research program

Electronic health records (EHRs) contain rich temporal data about respiratory viral infections, but methods to identify these infections from EHR data vary widely and lack robust validation. We developed computable phenotypes by integrating virus-specific International Classification of Diseases (ICD) billing codes, prescriptions, and laboratory results within 90-day episodes. Analysis of 265,222 participants with EHR data from the All of Us Research Program yielded national cohorts of varied size: large cohorts for SARS-CoV-2 (n = 28,729) and influenza (n = 19,784); medium cohorts for rhinovirus, human coronavirus, and respiratory syncytial virus (n = 1,161-1,620); and smaller cohorts for the other viruses (n = 238–486). Using laboratory results as a reference standard, phenotypes using virus-specific ICD codes and medications had variable sensitivity (8–67%) but high positive predictive value (PPV, 90–97%) for most viruses, while influenza virus and SARS-CoV-2 phenotypes had lower PPV (69–70%) that improved with the inclusion of additional ICD codes. Identified infections exhibited expected seasonal patterns matching CDC data. This integrated approach identified infections more effectively than individual components alone and demonstrated utility for severe infections in hospital settings. This method enables large-scale studies of host genetics, health disparities, and clinical outcomes across episodic diseases, with flexibility to optimize sensitivity or PPV depending on the specific research question.

Measuring the Impact of Data Quality and Computable Phenotypes on Potential Racial Disparities in Predicting Healthcare Utilization Among Type 2 Diabetes Populations.

Type 2 diabetes (T2D) computable phenotypes identify different denominator populations for downstream tasks. Differences in racial composition could introduce bias and lead to disparate disease management. The objective of this study was to assess potential racial disparities in predicting T2D healthcare utilization introduced by data quality and computable phenotypes. Four published and one local T2D phenotypes were applied to the EHR and claims datasets of a large academic medical center. Population characteristics were compared across phenotypes, stratified by race. We induced data incompleteness, inaccuracy, and untimeliness to measure the impact on denominator racial composition. We trained logistic classification models on each of the phenotype-specific populations separately and compared disparities in utilization prediction (i.e., inpatients (IP) and emergency room (ER) admissions). Model performance, such as mean AUC and positive/negative predictive values, were compared across phenotypes, stratified by race. Different T2D computable phenotypes identified populations with modestly different racial compositions. Black T2D patients had the highest average admissions to ER compared to other racial groups. Induced data quality challenges diminished patient counts across all racial groups proportionally. Charlson comorbidity score had the highest odds ratio in predicting IP and ER admissions across phenotypes and race groups. Specific T2D phenotypes showed the highest and lowest mean AUCs in predicting IP and ER admissions in Black and White populations; however, such results were not observed among Asian/Other populations. Utilization prediction differed among phenotypes and race groups. Understanding the complexities behind phenotypes, data quality, and predictive models could mitigate health disparity further downstream and inform clinical research and disease management.

Unlocking the Potential of Wear Time of a Wearable Device to Enhance Postpartum Depression Screening and Detection: Cross-Sectional Study

BackgroundPostpartum depression (PPD) is a mood disorder affecting 1 in 7 women after childbirth that is often underscreened and underdetected. If not diagnosed and treated, PPD is associated with long-term developmental challenges in the child and maternal morbidity. Wearable technologies, such as smartwatches and fitness trackers (eg, Fitbit), offer continuous and longitudinal digital phenotyping for mood disorder diagnosis and monitoring, with device wear time being an important yet understudied aspect.ObjectiveWe aimed to suggest that wear time of a wearable device may provide additional information about perinatal mental health to facilitate screening and early detection of PPD. We proposed that wear time of a wearable device may also be valuable for managing other mental health disorders.MethodsUsing the All of Us Research Program dataset, we identified females who experienced childbirth with and without PPD using computational phenotyping. We compared the percentage of days and number of hours per day females with and without PPD wore Fitbit devices during prepregnancy, pregnancy, postpartum, and PPD periods, determined by electronic health records. Comparisons between females with and without PPD were conducted using linear regression models. We also assessed the correlation between Fitbit wear time consistency (measured as the maximum number of consecutive days the Fitbit was worn) during prepregnancy and PPD periods in females with and without PPD using the Pearson correlation. All analyses were run with Bonferroni correction.ResultsOur findings showed a strong trend, although nonsignificant after multiple testing correction, that females in the PPD cohort wore their Fitbits more than those in non-PPD cohort during the postpartum (PPD cohort: mean 69.9%, 95% CI 42.7%-97%; non-PPD cohort: mean 50%, 95% CI 25.5%-74.4%; P=.02) and PPD periods (PPD cohort: mean 66.6%, 95% CI 37.9%-95.3%; non-PPD cohort: mean 46.4%, 95% CI 20.5%-72.2%; P=.02). We found no difference in the number of hours per day females in the PPD and non-PPD cohorts wore their Fitbit during any period of pregnancy. Finally, there was no relationship between the consistency of Fitbit wear time during prepregnancy and PPD periods (r=–0.05, 95% CI –0.46 to 0.38; P=.84); however, there was a trend, though nonsignificant, in Fitbit wear time consistency among females without PPD (r=0.25, 95% CI –0.02 to 0.49; P=.07).ConclusionsWe hypothesize that increased Fitbit wear time among females with PPD may be attributed to hypervigilance, given the common co-occurrence of anxiety symptoms. Future studies should assess the link between PPD, hypervigilance, and wear time patterns. We envision that wear time patterns of a wearable device combined with digital biomarkers such as sleep and physical activity could enhance early PPD detection using machine learning by alerting clinicians to potential concerns and facilitating timely screenings, which may have implications for other mental health disorders.

1092 A Narrative Review of Artificial Intelligence/Machine Learning Methods in Pediatric Sleep

Abstract Introduction Artificial Intelligence/Machine Learning (AI/ML) methods have risen in capability and application since 2012. In pediatric sleep health, these approaches have the potential to improve the surveillance and detection of pediatric sleep problems and disorders, automated scoring of polysomnography data, computational phenotyping, and clinical predictive algorithms. Sleep-related AI/ML innovations are endless and becoming increasingly prevalent in adult-focused sleep medicine and research. However, this review outlines the current AI/ML methods that have been deployed among pediatric sleep populations. Methods A review of the literature was conducted using PubMed, Embase, Web of Science, and Scopus to obtain articles related to AI/ML and pediatric sleep search terms. The criterion for article inclusion and full text review included: English language, published after the year 2012, and topically related to the application of an AI/ML method in pediatric sleep. Results This search yielded 3585 articles, which was narrowed to 1968 for screening after removing duplicates in Covidence. After applying screening criteria, 60 articles met the criteria for full-text review. AI/ML applications at the intersection of pediatric sleep include the following: predictive algorithms (e.g., obstructive sleep apnea detection, infant sleep, children with neurodevelopmental disabilities and comorbid sleep problems, adolescents with comorbid mental health conditions), automated sleep stage and duration scoring for pediatric polysomnography and actigraphy, and computational phenotyping in children with autism spectrum disorder. However, there is still need for leveraging studies leveraging feasible natural language processing approaches, healthcare-based detection tools, translational AI/ML methods that can be deployed as community-based solutions, and population-level surveillance of pediatric sleep. Conclusion This narrative review provides an overview of the existing AI/ML methodology that has been deployed in pediatric sleep, offers suggestions for future applications across age groups and settings, and an equity-based framework for designing future research that will further the intersecting fields of pediatric sleep and AI/ML. Despite the current progress that has been made in deploying AI/ML, design and implementation gaps remain among historically underserved sleep care populations. Future studies that center equity and population-level pediatric sleep health are crucial for innovative insights to be garnered. Support (if any)

0510 Large Language Models Streamline the Identification of an Insomnia Phenotype in Electronic Health Records

Abstract Introduction Insomnia is a common disorder characterized by difficulty falling and/or staying asleep. Its association with other health conditions and impact on quality of life underscore the importance of studying it at the population level. However, documentation of insomnia in electronic health records (EHRs) is inconsistent, and the condition is likely underdiagnosed. This study proposes the use of Natural Language Processing (NLP) methods to address the under-documentation of insomnia in EHRs, evaluating the performance of two Large Language Models (LLMs) in detecting an insomnia phenotype in clinical notes. Methods Two corpora of de-identified clinical notes were utilized in this study: 237 from the publicly available MIMIC-III database and 777 from patients in the University of Kansas Health System (UKHS) with family medicine encounters. Insomnia identification criteria were adapted from diagnostic guidelines, and gold standard datasets were created by manually annotating each note and labeling them as “Insomnia” or “Not Insomnia.” Two open-source LLMs, Llama3-70B and Llama3-405B, were trained to detect insomnia in the notes using prompt engineering. Values for precision, recall, and F1 score were used to evaluate LLM performance against more traditional NLP models. The LLM trained on the MIMIC-III was used to classify the UKHS notes. Results Both LLMs outperformed the conventional NLP models in detecting insomnia in a set of 70 test notes. Llama-3-405B achieved the highest values for precision (95.8), recall (100.0), and F1 score (97.9), and made the fewest classification errors. Both Llama3-70B and Llama3-405B both had a higher binary classification accuracy than proportion of error-free explanations for the same set of test notes. Transportability to UKHS notes revealed a decrease in performance, but with reasonable results (Llama3-405B precision 73.5, recall 86.2, F1 score 79.4). Conclusion The LLMs used in this study not only performed better than more traditional NLP models when detecting an insomnia phenotype from notes in EHRs but also exhibited adherence to a set of rules and provided explanations for their classifications. These results support the potential of generative AI to advance the study of insomnia on a larger scale through the reliable and accurate identification of a computable phenotype for insomnia. Support (if any)

0529 Geographic and Demographic Distribution of an Insomnia Phenotype in the All of Us Research Program

Abstract Introduction Insomnia is common among adults, but individuals’ experiences with the condition may not be well-reflected in electronic health records (EHRs). This leads to inconsistencies when using these codes alone to study insomnia. A computable phenotype for insomnia incorporating diagnosis codes, drug codes, and self-reported insomnia status could help us better understand the condition at the population level. This study aimed to use both EHR data and survey responses to understand the prevalence, geographic distribution, and demographic correlates of insomnia in a large cohort of adults in the All of Us Research Program. Methods Data from 99,824 participants who responded to the Personal/Family Health History survey and provided EHR and genomic sequencing data were included. We used an algorithm to identify individuals meeting a stringent (≥2 insomnia condition codes and ≥1 drug code in the EHR) or a broad (≥1 condition codes, drug codes, or self-reported insomnia) definition of insomnia. We estimated the overall clinical prevalence for each definition of insomnia among all participants, and stratified by sex, race, ethnicity, and geographic location. Results From included participants, 29,972 (30%) met the broad definition for insomnia and 17,596 (17.6%) met the stringent definition. For both definitions, groups that were overrepresented among insomnia cases relative to their proportion of the sample population included males, those identifying with a sex other than male or female, Non-Hispanic individuals, those identifying with an ethnicity other than Hispanic or Non-Hispanic, and White individuals. Black individuals were overrepresented among insomnia cases when the stringent, but not broad definition was used. Among U.S. states, Kansas had the highest insomnia prevalence according to both definitions (55.71% for broad and 39.5% for stringent), and North Dakota had the lowest (5.5% and 0%, respectively). Conclusion Using the broad definition resulted in a much greater estimated insomnia prevalence, reinforcing the idea that insomnia is underdiagnosed and underreported in EHRs. However, both definitions produced similar demographic and geographic trends. Further research should involve validating the algorithm used in this study to identify individuals with insomnia. Support (if any) All of Us Research Program is supported by the NIH, Office of the Director.

COMPAC: COMputable Phenotype for Asthma in Children.

Background Pediatric asthma is one of the most common chronic diseases of childhood. Reliable identification of pediatric asthma patients in electronic health records (EHRs) is essential for both research and clinical care. However, existing computable phenotypes (CPs) exhibit varying effectiveness. This study aims to evaluate current CPs and develop a new CP, named COMPAC (COMputable Phenotype for Asthma in Children), to improve EHR-based identification of pediatric asthma patients. Methods Multiple CP rules were designed using various combinations of diagnosis codes, prescriptions, and clinical note text. A cohort from the University of Florida Integrated Data Repository (IDR) was used for validation through manual chart reviews. Performance was assessed using standard metrics and compared to existing CPs. Additionally, bootstrapping and demographic subgroup analyses were conducted to compare the performance of the new COMPAC to previously published CPs. Results COMPAC demonstrated improved case identification compared to existing CPs, with high sensitivity (0.728; 95% confidence interval [CI]: 0.607-0.864), positive predictive value (0.886; 95% CI: 0.737-1.0), and an overall F1 score of 0.797 (95% CI: 0.682-0.90). Notably, COMPAC outperformed two previously published CPs in terms of F1 score. Performance varied across demographic subgroups, with COMPAC showing the best results in males, non-Hispanic Whites, and the 6-12 year-old age group, though its performance was lower in the 2-5 year-old age range. Conclusion COMPAC offers an improved approach for pediatric asthma case identification in EHRs. However, further validation across different sites and refinement to capture a broader range of clinical presentations are necessary to optimize its sensitivity and specificity.

Development and Validation of a Computable Radiation Therapy Phenotype.

Development and Validation of a Computable Radiation Therapy Phenotype.

Quantitative CT and Computational Fluid Dynamics-based Machine Learning Phenotypes of Lung Structure-function Abnormality

Quantitative CT and Computational Fluid Dynamics-based Machine Learning Phenotypes of Lung Structure-function Abnormality