AI-assisted screening for mild cognitive impairment using routine EHR data: a Gradient Boosting approach.

Introduction Ensuring equitable performance of sleep-related machine learning (ML) models is vital for public health and health equity. This research sought to determine the impact of sociodemographic and health disparities factors on the performance and characteristics of ML models designed to predict OSA treatment initiation. Methods Our data source was the All-Payer Claims Database (APCD) for the Wisconsin Health Information Organization (WHIO). Inclusion criteria included continuous insurance coverage for >12 months prior and >30 months after OSA diagnosis (defined by OCD code G47.33), and having undergone OSA diagnostic testing (defined by CPT codes). OSA treatment was defined based on durable medical equipment charges for PAP machines and supplies. Sociodemographic variables were extracted from APCD and included race, gender, age, and area socioeconomic deprivation (Area Deprivation Index; ADI). The ADI is a validated marker of health risk based on 17 health disparities factors ranking relative disadvantage across communities. Random Forest ML models were trained to predict OSA treatment initiation using sociodemographic variables and a medication history including 39,712 unique medications codified across 94 medication categories. Results Of N=6,026,463 subjects in the ACPD, n=154,821 underwent OSA diagnostic testing, and n=43,601 were diagnosed with OSA. 10-Fold Cross-Validation training-testing was applied to estimate sensitivity-specificity of ML models for predicting treatment initiation. Receiver operating characteristic curve area-under-the-curve (ROC-AUC) analyses were used to compare relative differences in predictive power of each variable. In ROC-AUC analysis of individual variables, the power for predicting OSA treatment were observed in relative rank order by age (0.568), race (0.547), ADI national-level (0.545), ADI state-level (0.544), gender (0.524), and medication history (0.516). In ROC-AUC analysis of combination variables, the highest ML model performance was observed in the combination of only three-variables (age-gender-race, 0.600), while the combination of all-variables showed an ROC-AUC value of (0.594), and the resulting performance difference was not statistically significant based on comparison of the ROC-AUC measures. Conclusion Demographic and health disparity factors may play an important role in future development of predictive AI/ML models. Population sleep health data represents an important resource to identify and bridge care gaps, reduce sleep health disparities, and achieve health equity. Support (if any)

AimsHeart failure (HF) is a clinical syndrome with no definitive diagnostic tests. HF registries are often based on manual reviews of medical records of hospitalized HF patients identified using International Classification of Diseases (ICD) codes. However, most HF patients are not hospitalized, and manual review of big electronic health record (EHR) data is not practical. The US Department of Veterans Affairs (VA) has the largest integrated healthcare system in the nation, and an estimated 1.5 million patients have ICD codes for HF (HF ICD‐code universe) in their VA EHR. The objective of our study was to develop artificial intelligence (AI) models to phenotype HF in these patients.Methods and resultsThe model development cohort (n = 20 000: training, 16 000; validation 2000; testing, 2000) included 10 000 patients with HF and 10 000 without HF who were matched by age, sex, race, inpatient/outpatient status, hospital, and encounter date (within 60 days). HF status was ascertained by manual chart reviews in VA's External Peer Review Program for HF (EPRP‐HF) and non‐HF status was ascertained by the absence of ICD codes for HF in VA EHR. Two clinicians annotated 1000 random snippets with HF‐related keywords and labelled 436 as HF, which was then used to train and test a natural language processing (NLP) model to classify HF (positive predictive value or PPV, 0.81; sensitivity, 0.77). A machine learning (ML) model using linear support vector machine architecture was trained and tested to classify HF using EPRP‐HF as cases (PPV, 0.86; sensitivity, 0.86). From the ‘HF ICD‐code universe’, we randomly selected 200 patients (gold standard cohort) and two clinicians manually adjudicated HF (gold standard HF) in 145 of those patients by chart reviews. We calculated NLP, ML, and NLP + ML scores and used weighted F scores to derive their optimal threshold values for HF classification, which resulted in PPVs of 0.83, 0.77, and 0.85 and sensitivities of 0.86, 0.88, and 0.83, respectively. HF patients classified by the NLP + ML model were characteristically and prognostically similar to those with gold standard HF. All three models performed better than ICD code approaches: one principal hospital discharge diagnosis code for HF (PPV, 0.97; sensitivity, 0.21) or two primary outpatient encounter diagnosis codes for HF (PPV, 0.88; sensitivity, 0.54).ConclusionsThese findings suggest that NLP and ML models are efficient AI tools to phenotype HF in big EHR data to create contemporary HF registries for clinical studies of effectiveness, quality improvement, and hypothesis generation.

We previously developed machine learning (ML) models for predicting cycloplegic spherical equivalent refraction (SER) and myopia using non-cycloplegic data and following a standardized protocol (cycloplegia with 0.5% tropicamide and biometry using NIDEK A-scan), but the models’ performance may not be generalizable to other settings. This study evaluated the performance of ML models in an independent cohort using a different cycloplegic agent and biometer. Chinese students (N = 614) aged 8–13 years underwent autorefraction before and after cycloplegia with 0.5% tropicamide (n = 505) or 1% cyclopentolate (n = 109). Biometric measures were obtained using an IOLMaster 700 (n = 207) or Optical Biometer SW-9000 (n = 407). ML models were evaluated using R2, mean absolute error (MAE), sensitivity, specificity, and area under the ROC curve (AUC). The XGBoost model predicted cycloplegic SER very well (R2 = 0.95, MAE (SD) = 0.32 (0.30) D). Both ML models predicted myopia well (random forest: AUC 0.99, sensitivity 93.7%, specificity 96.4%; XGBoost: sensitivity 90.1%, specificity 96.8%) and accurately predicted the myopia rate (observed 62.9%; random forest: 60.6%; XGBoost: 58.8%) despite heterogeneous cycloplegia and biometry factors. In this independent cohort of students, XGBoost and random forest performed very well for predicting cycloplegic SER and myopia status using non-cycloplegic data. This external validation study demonstrated that ML may provide a useful tool for estimating cycloplegic SER and myopia prevalence with heterogeneous clinical parameters, and study in additional populations is warranted.

Introduction:Asthma exacerbations are frequent causes of pediatric hospital admissions. We sought to develop a time series algorithm to forecast next-day daily asthma hospitalizations.Methods:Daily hospitalizations for asthma were collected at Cincinnati Children’s from January 1, 2016, to December 31, 2023. We evaluated Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing (ETS), Prophet, and Ensemble models to forecast next-day asthma hospitalizations validated on 2023 data, considering varying historical training data lengths. Forecasts were calibrated to identify days exceeding a 5% high-risk threshold of historical totals and considered multiple validation years and years before and during the COVID-19 pandemic.Results:A total of 5,593 hospital admissions were recorded for asthma. Over 2,922 days, 166 days met the 5% high-risk threshold equating to 6 or more admissions. The Ensemble (Median Absolute Percentage Error (MAPE): 46.7%; Positive Predictive Value (PPV): 0.278; Negative Predictive Value (NPV): 0.942; Area Under the ROC Curve (AUC): 0.740; Sensitivity: 0.800; Specificity: 0.656) model achieved higher accuracy of high-risk days than ARIMA (MAPE: 46.5%; PPV: 0.278; NPV: 0.942; AUC: 0.709; Sensitivity: 0.760; Specificity: 0.571), ETS (MAPE: 47.2%; PPV: 0.222; NPV: 0.939; AUC: 0.711; Sensitivity: 0.800; Specificity: 0.668), and Prophet (MAPE: 48.9%; PPV: 0.444; NPV: 0.951; AUC: 0.732; Sensitivity: 0.680; Specificity: 0.741) models.Conclusions:Our Ensemble model of mean predictions from ARIMA, ETS, and Prophet models was the most accurate in forecasting future asthma hospitalizations. Integrating forecasting techniques with clinical operations could enable proactive prevention through enhanced population care management.

109 Background: Machine learning tools based on EHR data hold promise to help avoid unnecessary risks associated with lung cancer and its treatment. Additionally, molecular genetic profiling is becoming an integral tool for clinicians to individualize treatment for lung cancer. However, relatively few survival models have been built that integrate this data in individualized predictive models. Here, we combine real-world EHR and tumor sequencing data from the VA Precision Oncology Data Repository (PODR) to build accurate individualized survival predictions in newly-diagnosed NSCLC patients. Methods: We identified a cohort of 356 VA patients newly diagnosed with NSCLC for whom EHR, cancer registry, and targeted tumor sequencing data is available in PODR. We defined 41 features reflecting 15 baseline clinical and demographic characteristics from the EHR and registry, such as age, race, stage, histology, and therapy. We also defined features reflecting 206 clinically actionable somatic variants. We selected 5 important variants for inclusion in the model, as well as the total number of mutations. We trained a random forests algorithm to predict 1-year survival. Precision, recall, and area under the ROC curve (AUC) were assessed using 5-fold cross validation. Results: Mean age at diagnosis was 66 years. The majority of patients had late stage disease (15% stage I, 6% II, 15% III, 44% IV), and 59% of patients received systemic therapy. 45% died within 1 year of diagnosis, and 55% survived past 1 year. Our predictive model for 1-year survival achieves strong results. Cross-validated AUC is 0.79 (SD 0.08), precision is 0.79 (SD 0.07), recall is 0.74 (SD 0.07), suggesting that the trained model combining clinical and genomic features is effective at predicting 1-year survival. Conclusions: By integrating real-world EHR and sequencing data, we built a highly accurate predictive model of 1-year survival in NSCLC patients at the VA. Such a model, after ongoing validation in a larger cohort, offers the ability to make individualized predictions that could inform patient care to improve outcomes.

BackgroundOpioids are a widely prescribed class of medication for pain management. However, they have variable efficacy and adverse effects among patients, due to the complex interplay between biological and clinical factors. Pharmacogenetic testing can be used to match patients’ genetic profiles to individualize opioid therapy, improving pain relief and reducing the risk of adverse effects. Despite its potential, the pharmacogenetic testing uptake (use of pharmacogenetic testing) remains low due to a range of barriers at the patient, health care provider, infrastructure, and financial levels. Since testing typically involves a shared decision between the provider and patient, predicting the likelihood of a patient undergoing pharmacogenetic testing and understanding the factors influencing that decision can help optimize resource use and improve outcomes in pain management.ObjectiveThis study aimed to develop machine learning (ML) models, identifying patients’ likelihood of pharmacogenetic uptake based on their demographics, clinical variables, medication use, and social determinants of health.MethodsWe used electronic health record data from a single center health care system to identify patients prescribed opioids. We extracted patients’ demographics, clinical variables, medication use, and social determinants of health, and developed and validated ML models, including a neural network, logistic regression, random forest, extreme gradient boosting (XGB), naïve Bayes, and support vector machines for pharmacogenetic testing uptake prediction based on procedure codes. We performed 5-fold cross-validation and created an ensemble probability-based classifier using the best-performing ML models for pharmacogenetic testing uptake prediction. Various performance metrics, uptake stratification analysis, and feature importance analysis were used to evaluate the performance of the models.ResultsThe ensemble model using XGB and support vector machine–radial basis function classifiers had the highest C-statistics at 79.61%, followed by XGB (78.94%), and neural network (78.05%). While XGB was the best-performing model, the ensemble model achieved a high accuracy (32,699/48,528, 67.38%), recall (537/702, 76.50%), specificity (32,162/47,826, 67.25%), and negative predictive value (32,162/32,327, 99.49%). The uptake stratification analysis using the ensemble model indicated that it can effectively distinguish across uptake probability deciles, where those in the higher strata are more likely to undergo pharmacogenetic testing in the real world (320/4853, 6.59% in the highest decile compared to 6/4853, 0.12% in the lowest). Furthermore, Shapley Additive Explanations value analysis using the XGB model indicated age, hypertension, and household income as the most influential factors for pharmacogenetic testing uptake prediction.ConclusionsThe proposed ensemble model demonstrated a high performance in pharmacogenetic testing uptake prediction among patients using opioids for pain. This model can be used as a decision support tool, assisting clinicians in identifying patients’ likelihood of pharmacogenetic testing uptake and guiding appropriate decision-making.

CT-DERIVED BODY COMPOSITION AND MACHINE LEARNING FOR PROGNOSTIC IMPROVEMENT IN COLORECTAL CANCER

To determine the most appropriate nutritional assessment tool for predicting the occurrence of major adverse cardiovascular events (MACE) within 1 year in elderly ACS patients undergoing PCI from four nutritional assessment tools including PNI, GNRI, CONUT, and BMI. Consecutive cases diagnosed with acute coronary syndrome (ACS) and underwent percutaneous coronary intervention (PCI) in the Department of Cardiovascular Medicine of the Air force characteristic medical center from 1 January 2020 to 1 April 2022 were retrospectively collected. The basic clinical characteristics and relevant test and examination indexes were collected uniformly, and the cases were divided into the MACE group (174 cases) and the non-MACE group (372 cases) according to whether a major adverse cardiovascular event (MACE) had occurred within 1 year. Predictive models were constructed to assess the nutritional status of patients with the Prognostic Nutritional Index (PNI), Geriatric Nutritional Risk Index (GNRI), Controlling nutritional status (CONUT) scores, and Body Mass Index (BMI), respectively, and to analyze their relationship with prognosis. The incremental value of the four nutritional assessment tools in predicting risk was compared using the Integrated Discriminant Improvement (IDI) and the net reclassification improvement (NRI). The predictive effect of each model on the occurrence of major adverse cardiovascular events (MACE) within 1 year in elderly ACS patients undergoing PCI was assessed using area under the ROC curve (AUC), calibration curves, decision analysis curves, and clinical impact curves; comparative analyses were performed. Among the four nutritional assessment tools, the area under the curve (AUC) was significantly higher for the PNI (AUC: 0.798, 95%CI 0.755–0.840 P < 0.001) and GNRI (AUC: 0.760, 95%CI 0.715–0.804 P < 0.001) than for the CONUT (AUC: 0.719,95%CI 0.673–0.765 P < 0.001) and BMI (AUC: 0.576, 95%CI 0.522–0.630 P < 0.001). The positive predictive value (PPV) of PNI: 67.67% was better than GNRI, CONUT, and BMI, and the negative predictive value (NPV): of 83.90% was better than CONUT and BMI and similar to the NPV of GNRI. The PNI, GNRI, and CONUT were compared with BMI, respectively. The PNI had the most significant improvement in the Integrated Discriminant Improvement Index (IDI) (IDI: 0.1732, P < 0.001); the PNI also had the most significant improvement in the Net Reclassification Index (NRI) (NRI: 0.8185, P < 0.001). In addition, of the four nutritional assessment tools used in this study, the PNI was more appropriate for predicting the occurrence of major adverse cardiovascular events (MACE) within 1 year in elderly ACS patients undergoing PCI.

Introduction: Arthritis is one of the most common chronic diseases. Early detection of arthritis and its progression can facilitate early intervention measures, lowering disease severity in patients. As electronic health records (EHR) become more accessible, this study assesses whether general health information and arthritis-related questionnaires can be used in arthritis diagnosis, without the involvement of costly imaging methods. Therefore, we created deep learning (DL) and machine learning (ML) models to explore the feasibility of combining EHR and modern computational tools to diagnose arthritis. Methods: A total of 782 arthritis patients and 4014 control patients were identified from the Osteoarthritis Initiative (OAI) – a ten-year-long observational study that included patient EHR in five time points. Six hundred variables were filtered by random forest classifier followed by manual filtering. Data were split properly to training, testing and validation set, and the training set was balanced. Sequential, nonsequential DL models, and five independent DL models for each time points were used. The accuracy, positive prevalence value (PPV), negative prevalence value (NPV), and area under curve (AUC), were assessed and compared with four classical ML models. SHAP (SHapley Additive exPlanations) summary analysis was also conducted. Results: Sequential and non-sequential deep learning models showed accuracies of ~ 0.97, and the four classical machine learning approaches showed accuracies of above 0.9. High positive and negative predicted values (&gt; 0.90) for all of the models suggested the potential clinical applicability of the model, while the SHAP analysis demonstrated its interpretability. Discussion: We tested various models and showed the ability to use machine learning methods for early diagnosis of arthritis with EHR. The models can be used as a screening tool to select susceptible patients for confirmatory tests such as X-ray and MRI. Identification of early disease states could facilitate protective measures that slow disease progression.

Machine learning based survival prediction of DLBCL: Using multimodal data.

Metabolic dysfunction-associated steatotic liver disease (MASLD) can progress to metabolic dysfunction-associated steatohepatitis (MASH) and liver fibrosis, contributing to a heavier global health burden. Non-invasive diagnostic tools developed using machine learning (ML) and deep learning (DL), two representative artificial intelligence algorithms, are increasingly being explored for MASH and its related fibrosis assessment. This study aimed to compare the diagnostic performance of different ML and DL models and identify the top-performing models for diagnosing MASH and associated liver fibrosis. A systematic review and meta-analysis were conducted across PubMed, Web of Science, Embase and Cochrane Library from inception to May 18, 2025. Pooled area under the receiver operator characteristic curve (AUROC) values with 95 % confidence interval (CI) were calculated. Accuracy, specificity, sensitivity, positive predictive values, and negative predictive values were also recorded. Of 4,314 studies initially identified, 106 met the inclusion criteria, with 35 studies (ML: n=28; DL: n=7) providing data for analysis. Logistic Regression and Neural Network are the most commonly algorithms applied in ML and DL, respectively. The pooled AUROCs for diagnosing MASH were 0.833 (95 %CI: 0.806-0.860) for ML models and 0.841 (95 %CI: 0.782-0.900) for DL models. Light Gradient Boosting Machine (LightGBM) and ResNet50 were the best-performing models for diagnosing MASH within ML and DL algorithms, respectively, achieving corresponding AUROCs of 0.920 (95 %CI: 0.916-0.924) and 0.960 (95 %CI: 0.951-0.969). For fibrosis diagnosis, ML models had a pooled AUROC of 0.826 (95 %CI: 0.792-0.860), with Categorical Boosting (CatBoost) achieving the highest AUROC of 0.960 (95 %CI: 0.950-0.970). DL models yielded the pooled AUROC of 0.875 (95 %CI: 0.816-0.934) for fibrosis diagnosis. Both ML and DL models demonstrated strong diagnostic performance for MASH and liver fibrosis, with DL achieving marginally higher AUROCs. AI-driven approaches show promise in MASLD management.

Urinary tract infections (UTIs) are a major cause of emergency hospital admissions, but it remains challenging to diagnose them reliably. Application of machine learning (ML) to routine patient data could support clinical decision-making. We developed a ML model predicting bacteriuria in the ED and evaluated its performance in key patient groups to determine scope for its future use to improve UTI diagnosis and thus guide antibiotic prescribing decisions in clinical practice. We used retrospective electronic health records from a large UK hospital (2011-2019). Non-pregnant adults who attended the ED and had a urine sample cultured were eligible for inclusion. The primary outcome was predominant bacterial growth ≥104 cfu/mL in urine. Predictors included demography, medical history, ED diagnoses, blood tests, and urine flow cytometry. Linear and tree-based models were trained via repeated cross-validation, re-calibrated, and validated on data from 2018/19. Changes in performance were investigated by age, sex, ethnicity, and suspected ED diagnosis, and compared to clinical judgement. Among 12,680 included samples, 4,677 (36.9%) showed bacterial growth. Relying primarily on flow cytometry parameters, our best model achieved an area under the ROC curve (AUC) of 0.813 (95% CI 0.792-0.834) in the test data, and achieved both higher sensitivity and specificity compared to proxies of clinician's judgement. Performance remained stable for white and non-white patients but was lower during a period of laboratory procedure change in 2015, in patients ≥65 years (AUC 0.783, 95% CI 0.752-0.815), and in men (AUC 0.758, 95% CI 0.717-0.798). Performance was also slightly reduced in patients with recorded suspicion of UTI (AUC 0.797, 95% CI 0.765-0.828). Our results suggest scope for use of ML to inform antibiotic prescribing decisions by improving diagnosis of suspected UTI in the ED, but performance varied with patient characteristics. Clinical utility of predictive models for UTI is therefore likely to differ for important patient subgroups including women <65 years, women ≥65 years, and men. Tailored models and decision thresholds may be required that account for differences in achievable performance, background incidence, and risks of infectious complications in these groups.

A machine learning (ML) model improved the prediction of pulmonary hypertension due to left heart disease (PH-LHD) in suspected pulmonary arterial hypertension (PAH) subjects using non-invasive predictors. (<i>K. Swinnen et al.</i>, ERS 2021) We aimed to externally validate this model to assess its generalizability to different populations. The ML model was developed using non-invasive predictors from 344 subjects from the University Hospitals of Leuven; demographics, medical history, echocardiographic, ECG and lung function parameters, and lab results were extracted from patient files. During internal validation (n=104), the ML model outperformed the Jacobs score (<i>W. Jacobs et al.</i>, ERJ 2015) with sensitivity 70%, specificity 100%, positive predictive value (PPV) 100% and negative predictive value (NPV) 78%. The external validation was performed in a cohort of 165 subjects (91 PAH and 74 PH-LHD) from Erasme Hospital Brussels. The ML model achieved similar performances (sensitivity 64%, specificity 100%, PPV 100% and NPV 78%) during external validation. Of the 74 PH-LHD subjects, 47 were correctly identified non-invasively without false positives <i>versus</i> 2 with the Jacobs score. The cohorts had significant differences in baseline characteristics, including important parameters contributing to the model predictions. Despite differences between populations, the ML model showed good generalizability when tested on subjects from a different centre with only a slight decrease in sensitivity. This model is ready to be tested in clinical practice and could substantially reduce the number of right heart catheterizations in patients with high probability for PH-LHD.

Hospital readmission after third-stage palliation for single ventricle physiology (Fontan cavopulmonary anastomosis) approaches 25%. The cause for readmissions is varied, and there is no clear way to predict those at risk for readmission. If an effective predictive model can be used, physicians can preempt readmission in either the inpatient or outpatient setting. In this study, we incorporated electronic health records (EHRs), such as vital signs, administered medications, laboratory results, and demographic information, to determine whether a patient will require readmission after undergoing Fontan surgery. The EHR data were extracted for 338 patients admitted to the Children’s Healthcare of Atlanta from May 2009 to August 2020 for Fontan surgery. 38.8% of the patients were readmitted. Machine learning models, such as extreme gradient boosting, random forest, decision tree, and logistic regression, were employed to make these predictions. The random forest classifier outperformed the rest of the models and achieved 89.6% accuracy [82.1%, 95.5%], 77.8% sensitivity [63.0%, 92.6%], 97.5% specificity [90%, 100%], 95.2% positive predictive value [83.7%, 100%], 86.7% negative predictive value [78.7%, 94.9%], 82.2% area under precision-recall curve [69.8%, 92.7%], 87.0% area under receiver operating characteristic [78.4%, 94.4%], and 85.1% F1 score [72.7%, 94.1%].

IntroductionYoung children commonly wheeze but only some have asthma later in life. Asthma prediction tools have poor predictive performance and few have been validated. We aimed to develop a robust...