Predicting Diagnostic Progression to Schizophrenia or Bipolar Disorder via Machine Learning

PurposeThe purpose of this study was to develop models that predict which patients with glaucoma will progress to require surgery, combining structured data from electronic health records (EHRs) and retinal fiber layer optical coherence tomography (RNFL OCT) scans.MethodsEHR data (demographics and clinical eye examinations) and RNFL OCT scans were identified for patients with glaucoma from an academic center (2008–2023). Comparing the novel TabNet deep learning architecture to a baseline XGBoost model, we trained and evaluated single modality models using either EHR or RNFL features, as well as fusion models combining both EHR and RNFL features as inputs, to predict glaucoma surgery within 12 months (binary).ResultsWe had 1472 patients with glaucoma who were included in this study, of which 29.9% (N = 367) progressed to glaucoma surgery. The TabNet fusion model achieved the highest performance on the test set with an area under the receiver operating characteristic curve (AUROC) of 0.832, compared to the XGBoost fusion model (AUROC = 0.747). EHR only models performed with an AUROC of 0.764 and 0.720 for the deep learning model and XGBoost models, respectively. RNFL only models performed with an AUROC of 0.624 and 0.633 for the deep learning and XGBoost models, respectively.ConclusionsFusion models which integrate both RNFL with EHR data outperform models only utilizing one datatype or the other to predict glaucoma progression. The deep learning TabNet architecture demonstrated superior performance to traditional XGBoost models.Translational RelevancePrediction models that utilize the wealth of structured clinical and imaging data to predict glaucoma progression could form the basis of future clinical decision support tools to personalize glaucoma care.

The screening and diagnosis of osteoporosis in patients with type 2 diabetes mellitus (T2DM) based on bone mineral density remains challenging because of the limited availability and accessibility of dual-energy X-ray absorptiometry. We aimed to develop and validate models to predict the risk of osteoporosis in postmenopausal women with T2DM based on machine learning (ML) algorithms. This retrospective study included 303 postmenopausal women with T2DM. To develop prediction models for osteoporosis, we applied nine ML algorithms combined with demographic, clinical, and laboratory data. The least absolute shrinkage and selection operator were used to perform feature selection. We used the bootstrap resampling technique for model training and validation. To test the performance of the models, we calculated indices including the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, calibration curve, and decision curve analysis. Furthermore, we conducted fivefold cross-validation for parameter optimization and model validation. Feature importance was assessed using the SHapley additive explanation (SHAP). We identified 10 independent predictors as the most valuable features. An AUROC of 0.616-1.000 was observed for nine ML algorithms. The extreme gradient boosting (XGBoost) model exhibited the best performance, outperforming conventional risk assessment tools and registering 0.993 in the training set, 0.798 in the validation set, and 0.786 in the test set for fivefold cross-validation. Using SHAP, we found that the explanatory variables contributed to the model and their relationship with osteoporosis occurrence. Furthermore, we developed a user-friendly tool for calculating the risk of osteoporosis. With the integration of demographic and clinical risk factors, ML algorithms can accurately predict osteoporosis. The XGBoost model showed ideal performance. With the incorporation of these models in the clinic, patients may benefit from early osteoporosis diagnosis and treatment.

Patients with non-small cell lung cancer (NSCLC) and patients with NSCLC combined with chronic obstructive pulmonary disease (COPD) have similar physiological conditions in early stages, and the latter have shorter survival times and higher mortality rates. The purpose of this study was to develop and compare machine learning models to identify future diagnoses of COPD combined with NSCLC patients based on the patient's disease and routine clinical data. Data were obtained from 237 patients with COPD combined with NSCLC as well as NSCLC admitted to Ningxia Hui Autonomous Region People's Hospital from October 2013 to July 2022. Six machine learning algorithms (K-nearest neighbor, logistic regression, eXtreme gradient boosting, support vector machine, naïve Bayes, and artificial neural network) were used to develop prediction models for NSCLC combined with COPD. Sensitivity, specificity, positive predictive value, negative predictive value, accuracy, F1 score, Mathews correlation coefficient (MCC), Kappa, area under the receiver operating characteristic curve (AUROC)and area under the precision-recall curve (AUPRC) were used as performance indicators to evaluate the performance of the models. 135 patients with NSCLC combined with COPD, 102 patients with NSCLC were included in the study. The results showed that pulmonary function and emphysema were important risk factors and that the support vector machine-based identification model showed optimal performance with accuracy:0.946, recall:0.940, specificity:0.955, precision:0.972, npv:0.920, F1 score:0.954, MCC:0.893, Kappa:0.888, AUROC:0.975, AUPRC:0.987. The use of machine learning tools combining clinical symptoms and routine examination data features is suitable for identifying the risk of concurrent NSCLC in COPD patients.

BackgroundClinical determination of patients at high risk of poor surgical outcomes is complex and may be supported by clinical tools to summarize the patient’s own personalized electronic health record (EHR) history and vitals data through predictive risk models. Since prior models were not readily available for EHR-integration, our objective was to develop and validate a risk stratification tool, named the Assessment of Geriatric Emergency Surgery (AGES) score, predicting risk of 30-day major postoperative complications in geriatric patients under consideration for urgent and emergency surgery using pre-surgical existing electronic health record (EHR) data.MethodsPatients 65-years and older undergoing urgent or emergency non-cardiac surgery within 21 hospitals 2017–2021 were used to develop the model (randomly split: 80% training, 20% test). The primary outcome was a 30-day composite outcome including several postoperative major complications and mortality; secondary outcomes included common individual complications included in the primary composite outcome (sepsis, progressive renal insufficiency or renal failure, and mortality). Patients’ EHR-based clinical history, vital signs, labs, and demographics were included in logistic regression, LASSO, decision tree, Random Forest, and XGBoost models. Area under the receiver operating characteristics curve (AUCROC) was used to compare model performance.ResultsOverall, 66,262 patients (median [IQR] age 78 [(70.9–84.0], female 53.9%, White race 68.5%) received urgent or emergency non-cardiac surgery (25.7% orthopedic cases, 21.9% general surgery cases). AUCROC ranged from 0.655 (Decision Tree) – 0.804 (XGBoost) for the primary composite outcome. XGBoost AUCROC was 0.823, 0.781, and 0.839 in predicting outcomes of sepsis, progressive renal insufficiency or renal failure, and mortality, respectively.ConclusionsWe developed a model to accurately predict major postoperative complications in geriatric patients undergoing urgent or emergency surgery using the patient’s own existing EHR data. EHR implementation of this model could efficiently support clinicians’ surgical risk assessment and perioperative decision-making discussions in this vulnerable patient population.

BACKGROUND Postoperative delirium (POD) is a common complication after major surgery and is associated with poor outcomes in older adults. Early identification of patients at high risk of POD can enable targeted prevention efforts. However, existing POD prediction models require inpatient data collected during the hospital stay, which delays predictions and limits scalability. OBJECTIVE This study aimed to develop and externally validate a machine learning-based prediction model for POD using routine electronic health record (EHR) data. METHODS We identified all surgical encounters from 2014 to 2021 for patients aged 50 years and older who underwent an operation requiring general anesthesia, with a length of stay of at least 1 day at 3 Indiana hospitals. Patients with preexisting dementia or mild cognitive impairment were excluded. POD was identified using Confusion Assessment Method records and delirium International Classification of Diseases (ICD) codes. Controls without delirium or nurse-documented confusion were matched to cases by age, sex, race, and year of admission. We trained logistic regression, random forest, extreme gradient boosting (XGB), and neural network models to predict POD using 143 features derived from routine EHR data available at the time of hospital admission. Separate models were developed for each hospital using surveillance periods of 3 months, 6 months, and 1 year before admission. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC). Each model was internally validated using holdout data and externally validated using data from the other 2 hospitals. Calibration was assessed using calibration curves. RESULTS The study cohort included 7167 delirium cases and 7167 matched controls. XGB outperformed all other classifiers. AUROCs were highest for XGB models trained on 12 months of preadmission data. The best-performing XGB model achieved a mean AUROC of 0.79 (SD 0.01) on the holdout set, which decreased to 0.69-0.74 (SD 0.02) when externally validated on data from other hospitals. CONCLUSIONS Our routine EHR-based POD prediction models demonstrated good predictive ability using a limited set of preadmission and surgical variables, though their generalizability was limited. The proposed models could be used as a scalable, automated screening tool to identify patients at high risk of POD at the time of hospital admission.

BackgroundFor the clinical care of patients with well-established diseases, randomized trials, literature, and research are supplemented with clinical judgment to understand disease prognosis and inform treatment choices. In the void created by a lack of clinical experience with COVID-19, artificial intelligence (AI) may be an important tool to bolster clinical judgment and decision making. However, a lack of clinical data restricts the design and development of such AI tools, particularly in preparation for an impending crisis or pandemic.ObjectiveThis study aimed to develop and test the feasibility of a “patients-like-me” framework to predict the deterioration of patients with COVID-19 using a retrospective cohort of patients with similar respiratory diseases.MethodsOur framework used COVID-19–like cohorts to design and train AI models that were then validated on the COVID-19 population. The COVID-19–like cohorts included patients diagnosed with bacterial pneumonia, viral pneumonia, unspecified pneumonia, influenza, and acute respiratory distress syndrome (ARDS) at an academic medical center from 2008 to 2019. In total, 15 training cohorts were created using different combinations of the COVID-19–like cohorts with the ARDS cohort for exploratory purposes. In this study, two machine learning models were developed: one to predict invasive mechanical ventilation (IMV) within 48 hours for each hospitalized day, and one to predict all-cause mortality at the time of admission. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value, and negative predictive value. We established model interpretability by calculating SHapley Additive exPlanations (SHAP) scores to identify important features.ResultsCompared to the COVID-19–like cohorts (n=16,509), the patients hospitalized with COVID-19 (n=159) were significantly younger, with a higher proportion of patients of Hispanic ethnicity, a lower proportion of patients with smoking history, and fewer patients with comorbidities (P<.001). Patients with COVID-19 had a lower IMV rate (15.1 versus 23.2, P=.02) and shorter time to IMV (2.9 versus 4.1 days, P<.001) compared to the COVID-19–like patients. In the COVID-19–like training data, the top models achieved excellent performance (AUROC>0.90). Validating in the COVID-19 cohort, the top-performing model for predicting IMV was the XGBoost model (AUROC=0.826) trained on the viral pneumonia cohort. Similarly, the XGBoost model trained on all 4 COVID-19–like cohorts without ARDS achieved the best performance (AUROC=0.928) in predicting mortality. Important predictors included demographic information (age), vital signs (oxygen saturation), and laboratory values (white blood cell count, cardiac troponin, albumin, etc). Our models had class imbalance, which resulted in high negative predictive values and low positive predictive values.ConclusionsWe provided a feasible framework for modeling patient deterioration using existing data and AI technology to address data limitations during the onset of a novel, rapidly changing pandemic.

Background/Objectives: Early detection of postnatal growth failure (PGF) is essential for optimizing nutritional management in preterm infants, as PGF is associated with adverse neurodevelopmental outcomes. Early prediction remains difficult because postnatal growth is influenced by multiple clinical factors including gestation age, birth weight, nutritional status, and comorbidities. Machine-learning approaches have been proposed to predict complex neonatal outcomes. This study compared the predictive performance of neonatologists with that of a machine-learning model for predicting PGF. Methods: PGF was defined as a decrease in weight z-score greater than 1.28 at discharge compared with birth. A machine-learning model based on extreme gradient boosting (XGBoost) was trained using a dataset of 7954 very low birth weight (VLBW) infants. Nine neonatologists independently assessed 100 clinical cases through a questionnaire-based evaluation, including 50 patients with PGF. Predictive performance was evaluated using seven metrics: area under the receiver operating characteristic curve (AUROC), accuracy, error rate, positive predictive value (PPV), sensitivity, specificity, and F1 score. Results: The neonatologists had a median of 5 years (range: 4–10 years) of clinical experience. The median prediction score among the neonatologists was 52/100 (range, 44–60), whereas the XGBoost model achieved 79/100. The XGBoost model achieved an AUROC of 0.79, accuracy of 0.79, error rate of 0.21, sensitivity of 0.82, and an F1 score of 0.80, demonstrating superior overall performance compared to the neonatologists. In addition, the XGBoost model had a lower error rate than the neonatologists (0.21 vs. 0.49), whereas specificity (0.76 vs. 0.86) and PPV (0.77 vs. 0.53) did not differ significantly. Conclusions: The machine-learning model demonstrated superior or comparable predictive performance to that of neonatologists in detecting PGF. Machine-learning-based prediction models may support early risk stratification and targeted nutritional management in VLBW infants.

Older adults with diabetes are at high risk of severe hypoglycemia (SH). Many machine-learning (ML) models predict short-term hypoglycemia are not specific for older adults and show poor precision-recall. We aimed to develop a multidimensional, electronic health record (EHR)-based ML model to predict one-year risk of SH requiring hospitalization in older adults with diabetes. We adopted a case-control design for a retrospective territory-wide cohort of 1,456,618 records from 364,863 unique older adults (age ≥65 years) with diabetes and at least 1 Hong Kong Hospital Authority attendance from 2013 to 2018. We used 258 predictors including demographics, admissions, diagnoses, medications, and routine laboratory tests in a one-year period to predict SH events requiring hospitalization in the following 12 months. The cohort was randomly split into training, testing, and internal validation sets in a 7:2:1 ratio. Six ML algorithms were evaluated including logistic-regression, random forest, gradient boost machine, deep neural network (DNN), XGBoost, and Rulefit. We tested our model in a temporal validation cohort in the Hong Kong Diabetes Register with predictors defined in 2018 and outcome events defined in 2019. Predictive performance was assessed using area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC) statistics, and positive predictive value (PPV). We identified 11,128 SH events requiring hospitalization during the observation periods. The XGBoost model yielded the best performance (AUROC = 0.978 [95% CI 0.972 to 0.984]; AUPRC = 0.670 [95% CI 0.652 to 0.688]; PPV = 0.721 [95% CI 0.703 to 0.739]). This was superior to an 11-variable conventional logistic-regression model comprised of age, sex, history of SH, hypertension, blood glucose, kidney function measurements, and use of oral glucose-lowering drugs (GLDs) (AUROC = 0.906; AUPRC = 0.085; PPV = 0.468). Top impactful predictors included non-use of lipid-regulating drugs, in-patient admission, urgent emergency triage, insulin use, and history of SH. External validation in the HKDR cohort yielded AUROC of 0.856 [95% CI 0.838 to 0.873]. Main limitations of this study included limited transportability of the model and lack of geographically independent validation. Our novel-ML model demonstrated good discrimination and high precision in predicting one-year risk of SH requiring hospitalization. This may be integrated into EHR decision support systems for preemptive intervention in older adults at highest risk.

Machine-learning-derived online prediction models of outcomes for patients with cholelithiasis-induced acute cholangitis: development and validation in two retrospective cohorts

Category:Ankle Arthritis; AnkleIntroduction/Purpose:Ankle arthrodesis and total ankle replacement are the most commonly performed procedures for surgical management of ankle arthritis. Arthrodesis provides effective pain relief but the rate of complications after arthrodesis is higher as it is more commonly performed in patients with comorbidities that preclude ankle replacement. Accurately risk- stratifying patients who undergo ankle arthrodesis would be of great utility, given the significant cost and morbidity associated with developing major perioperative complications. There is a paucity of accurate prediction models that can be used to pre- operatively risk-stratify patients for ankle arthrodesis. We aim to develop a machine learning (ML) algorithm for prediction of major perioperative complication after ankle arthrodesis as well as compare its performance against traditional predictive models based on logistic regression.Methods:This is a retrospective cohort study of adult patients who underwent ankle arthrodesis at any non-federal California hospital between 2015 and 2017. The primary outcome was readmission within 30 days or major perioperative complication - venous thromboembolism within 30 days, myocardial infarction within 7 days, pneumonia within 7 days, systemic infection within 7 days, surgical site bleeding within 90 days, and wound complications within 90 days. We build ML and logistic regression models that span different classes of modeling approaches: XGBoost, AdaBoost, Gradient Boosting, and Random Forest. Discrimination and calibration were assessed using area under the receiver operating characteristic curve (AUROC) and Brier score, respectively. We utilize a partial dependence function to measure the importance of an individual feature by assessing the average effect in predicted risks when its value is altered. We rank the contribution of the included variables to the prediction of adverse outcomes.Results:A total of 1,084 patients met inclusion criteria for this study. There were 131 major complications or readmission (12.1%). The optimized XGBoost algorithm demonstrates higher discrimination (AUROC: 0.707 + 0.052) compared to LR (0.691 + 0.055). The receiver operating characteristic curves for the XGBoost and logistic regression models are visualized in Figure 1. XGBoost also outperforms the three other ML models. This model was well calibrated (Brier score: 0.103 + 0.001). The variables most important for the XGBoost model include diabetes, chronic kidney disease, implant complication, and major fracture. Five of the ten most important features for XGBoost were markedly less important for the traditional logistic regression model: male sex, prior hip fracture, cardiorespiratory failure, acute renal failure, and dialysis status.Conclusion:We report a ML algorithm for prediction of major perioperative complications after ankle arthrodesis. The optimized XGBoost model is well-calibrated and demonstrates superior risk prediction to logistic regression. This tool may identify and address potentially modifiable risk factors, helping to accurately risk-stratify patients and decrease likelihood of major complications. Notably, the predictors most important for XGBoost are different from those for logistic regression. This suggests that the superior discriminative capability of ML methods stems from their ability to capture complex non-linear relationships between variables that logistic regression is unable to detect.

Parkinson disease (PD) is a rapidly growing neurodegenerative disorder that presents a significant public health challenge in aging societies, particularly in China. Although genetic, clinical, and environmental risk factors have been well-established, the predictive value of multidimensional factors (encompassing socioeconomic determinants, prodromal functional impairments, and lifestyle) remains underexplored in Chinese populations. We used data from the China Health and Retirement Longitudinal Study (n = 13,649; PD cases = 269) to develop and validate a machine learning prediction model for PD. Candidate predictors covered cognition, socioeconomic status, comorbidities, and lifestyle. Feature selection was performed using an ensemble approach (boruta, least absolute shrinkage and selection operator, and minimum redundancy maximum relevance), followed by extreme gradient boosting (XGBoost) model development with Bayesian hyperparameter optimization. Data were randomly split into a training set (70%) and a validation set (30%) for internal validation. Model performance was evaluated using area under the receiver operating characteristic curve, decision curve analysis, sensitivity, specificity, positive predictive value, negative predictive value, and F1-score. Model interpretability was evaluated using SHapley Additive exPlanations. Key predictors included cognitive impairment (memory and executive deficits), physical inactivity (low metabolic-equivalent scores), socioeconomic indicators (retirement status and financial support), educational attainment, and comorbidities (e.g., liver disease and wrist pain). In the validation set, the optimized XGBoost model achieved an area under the curve of 0.967 (95% confidence interval 0.962–0.972), sensitivity 82.0%, specificity 94.7%, positive predictive value 81.5%, negative predictive value 94.9%, and F1-score 81.7. SHapley Additive exPlanations analyses supported the consistency and relative importance of these predictors. An explainable XGBoost model integrating multidimensional national data showed excellent internal validation performance for identifying individuals with PD in China Health and Retirement Longitudinal Study. This tool may help support PD risk stratification in Chinese populations; external validation and prospective evaluation are warranted before clinical implementation.

Oxidative stress is related to the pathogenesis of mood disorders, and the level of oxidative stress may differ between bipolar disorder (BD) and major depressive disorder (MDD). This study aimed to detect the differences in non-enzymatic antioxidant levels between BD and MDD and assess the predictive values of non-enzymatic antioxidants in mood disorders by applying a machine learning model. Peripheral uric acid (UA), albumin (ALB), and total bilirubin (TBIL) were measured in 1,188 participants (discover cohort: 157 with BD and 544 with MDD; validation cohort: 119 with BD and 95 with MDD; 273 healthy controls). An extreme gradient boosting (XGBoost) model and a logistic regression model were used to assess the predictive effect. All three indices differed between patients with mood disorders and healthy controls; in addition, the levels of UA in patients with BD were higher than those of patients with MDD. After treatment, UA levels increased in the MDD group, while they decreased in the BD group. Finally, we entered age, sex, UA, ALB, and TBIL into the XGBoost model. The area under the curve (AUC) of the XGBoost model for distinguishing between BD and MDD reached 0.849 (accuracy = 0.808, 95% CI = 0.719-0.878) and for distinguishing between BD with depression episode (BD-D) and MDD was 0.899 (accuracy = 0.891, 95% CI = 0.856-0.919). The models were validated in the validation cohort. The most important feature distinguishing between BD and MDD was UA. Peripheral non-enzymatic antioxidants, especially the UA, might be a potential biomarker capable of distinguishing between BD and MDD.

Electronic health records (EHRs) with clinical decision support tools are now ubiquitous in healthcare organizations. Clinical foundation models (CFMs) pretrained on large-scale, heterogeneous structured EHR data have emerged as a powerful approach to improve predictive performance and generalizability. Meanwhile, large language models (LLMs) pretrained on broad data sources are being applied to an expanding range of healthcare tasks. However, it remains unclear whether generalist LLMs can match specialized CFMs for disease risk prediction using structured clinical data. We compared CFMs (Med-BERT, CLMBR) against fine-tuned generalist LLMs (Mistral, LLaMA-2/3/3.1), a clinical LLM (Me-LLaMA), and LLM-generated embeddings paired with simple classifiers (using DeepSeek, Qwen3, and GPT-OSS) on two disease risk prediction tasks: heart failure risk among diabetic patients (DHF) and pancreatic cancer diagnosis (PaCa). Evaluations spanned multi-site EHR data, claims data, and an open-source single-institution benchmark (EHRSHOT). Performance was assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). On larger EHR and claims cohorts (>30,000 patients), fine-tuned CFMs outperformed fine-tuned LLMs by a small but statistically significant margin (<1% AUROC). The clinical LLM performed comparably to generalist LLMs despite being smaller. On the open-source PaCa cohort (3,810 patients, 199 cases), LLMs achieved slightly higher AUROCs that were not statistically significant (LLaMA-3.1-70B 86.1% vs. Med-BERT 85.3%, p=0.27), but CFMs achieved significantly higher AUPRC (Med-BERT 55.9% vs. LLaMA-3.1-70B 41.1%, p=0.001). Notably, LLM-generated trajectory embeddings paired with logistic regression or a simple MLP, without any LLM fine-tuning, achieved the best overall performance, with AUROC exceeding 90% (Qwen3) and AUPRC reaching 66% (GPT-OSS 20B). LLM-generated embeddings with lightweight classifiers outperformed both fine-tuned CFMs and fine-tuned LLMs on AUROC and AUPRC. While these results demonstrate the potential of generalist models to match or surpass specialized CFMs, their substantially greater computational cost and variable AUPRC performance in the fine-tuning setting warrant caution. We provide a reproducible evaluation framework and codebase to support continued benchmarking.

Accurate hospital length of stay (LoS) prediction enables efficient resource management. Conventional LoS prediction models with limited covariates and nonstandardized data have limited reproducibility when applied to the general population. In this study, we developed and validated a machine learning (ML)-based LoS prediction model for planned admissions using the Observational Medical Outcomes Partnership Common Data Model (OMOP CDM). Retrospective patient-level prediction models used electronic health record (EHR) data converted to the OMOP CDM (version 5.3) from Seoul National University Bundang Hospital (SNUBH) in South Korea. The study included 137,437 hospital admission episodes between January 2016 and December 2020. Covariates from the patient, condition occurrence, medication, observation, measurement, procedure, and visit occurrence tables were included in the analysis. To perform feature selection, we applied Lasso regularization in the logistic regression. The primary outcome was an LoS of 7 days or longer, while the secondary outcome was an LoS of 3 days or longer. The prediction models were developed using 6 ML algorithms, with the training and test set split in a 7:3 ratio. The performance of each model was evaluated based on the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). Shapley Additive Explanations (SHAP) analysis measured feature importance, while calibration plots assessed the reliability of the prediction models. External validation of the developed models occurred at an independent institution, the Seoul National University Hospital. The final sample included 129,938 patient entry events in the planned admissions. The Extreme Gradient Boosting (XGB) model achieved the best performance in binary classification for predicting an LoS of 7 days or longer, with an AUROC of 0.891 (95% CI 0.887-0.894) and an AUPRC of 0.819 (95% CI 0.813-0.826) on the internal test set. The Light Gradient Boosting (LGB) model performed the best in the multiclassification for predicting an LoS of 3 days or more, with an AUROC of 0.901 (95% CI 0.898-0.904) and an AUPRC of 0.770 (95% CI 0.762-0.779). The most important features contributing to the models were the operation performed, frequency of previous outpatient visits, patient admission department, age, and day of admission. The RF model showed robust performance in the external validation set, achieving an AUROC of 0.804 (95% CI 0.802-0.807). The use of the OMOP CDM in predicting hospital LoS for planned admissions demonstrates promising predictive capabilities for stays of varying durations. It underscores the advantage of standardized data in achieving reproducible results. This approach should serve as a model for enhancing operational efficiency and patient care coordination across health care settings.

BackgroundPancreatic cancer (PANC) is often diagnosed at late stages due to the absence of specific early symptoms, resulting in one of the highest cancer mortality rates. While imaging modalities such as MRI and CT offer high diagnostic accuracy, their population-wide application is however impractical due to the cost. Electronic health records (EHRs) provide a routine, easily accessible, longitudinal and scalable data source for risk prediction, particularly for diseases with no specific symptom such as PANC.MethodWe introduce PANCDetect, a multimodal framework that leverages large language model (LLM)-derived embeddings of diagnoses, procedures, medications, and laboratory tests, and integrates these data modalities through a Transformer-based architecture. We train the model on MarketScan (≈250M patients), and validate it externally on additional large real-world EHR datasets of University of Michigan Precision Health, or UMPH data (n≈6M patients) and OneFlorida+ data (n≈26M patients). We then fine-tuned the general model on UMPH EHR data. We evaluated the performance of both models using metrics including area under the receiver operating characteristic curve (AUROC) and area under the precision-recall-gain curve (AUPRG). We assessed the top predictive features with integrated gradients (IG).ResultIn the MarketScan cohort, PANCDetect achieved an AUROC of 0.812 and AUPRG of 0.851 at the 6-month prediction window, and an AUROC of 0.735 and AUPRG of 0.629 for 60-month prediction, significantly outperforming CancerRiskNet. External validation on UMPH and OneFlorida+ demonstrated good generalizability, with 6-months AUROC scores of 0.711 and 0.793, respectively. Fine-tuning on UMPH with laboratory data further improved performance, reaching an AUROC of 0.927 and an AUPRG of 0.979 at 6 months. Even at the 60-month horizon, the refined PANCDetect model maintained strong performance, with an AUROC of 0.835 and AUPRG of 0.911. Attribution analysis highlighted type 2 diabetes, pancreatic diseases, personal and family cancer history as the most important risk factors.ConclusionPANCDetect is the state-of-the-art method integrating multimodal EHR data with LLM embeddings for accurate, interpretable, and generalizable early prediction of pancreatic cancer. This framework holds promise for precision screening of high-risk patients, with the potential to improve survival outcomes without increasing healthcare costs.