Development of machine learning models to predict posterior capsule rupture based on the EUREQUO registry.

BackgroundIn the pediatric intensive care unit (PICU), quantifying illness severity can be guided by risk models to enable timely identification and appropriate intervention. Logistic regression models, including the pediatric index of mortality 2 (PIM-2) and pediatric risk of mortality III (PRISM-III), produce a mortality risk score using data that are routinely available at PICU admission. Artificial neural networks (ANNs) outperform regression models in some medical fields.ObjectiveIn light of this potential, we aim to examine ANN performance, compared to that of logistic regression, for mortality risk estimation in the PICU.MethodsThe analyzed data set included patients from North American PICUs whose discharge diagnostic codes indicated evidence of infection and included the data used for the PIM-2 and PRISM-III calculations and their corresponding scores. We stratified the data set into training and test sets, with approximately equal mortality rates, in an effort to replicate real-world data. Data preprocessing included imputing missing data through simple substitution and normalizing data into binary variables using PRISM-III thresholds. A 2-layer ANN model was built to predict pediatric mortality, along with a simple logistic regression model for comparison. Both models used the same features required by PIM-2 and PRISM-III. Alternative ANN models using single-layer or unnormalized data were also evaluated. Model performance was compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision recall curve (AUPRC) and their empirical 95% CIs.ResultsData from 102,945 patients (including 4068 deaths) were included in the analysis. The highest performing ANN (AUROC 0.871, 95% CI 0.862-0.880; AUPRC 0.372, 95% CI 0.345-0.396) that used normalized data performed better than PIM-2 (AUROC 0.805, 95% CI 0.801-0.816; AUPRC 0.234, 95% CI 0.213-0.255) and PRISM-III (AUROC 0.844, 95% CI 0.841-0.855; AUPRC 0.348, 95% CI 0.322-0.367). The performance of this ANN was also significantly better than that of the logistic regression model (AUROC 0.862, 95% CI 0.852-0.872; AUPRC 0.329, 95% CI 0.304-0.351). The performance of the ANN that used unnormalized data (AUROC 0.865, 95% CI 0.856-0.874) was slightly inferior to our highest performing ANN; the single-layer ANN architecture performed poorly and was not investigated further.ConclusionsA simple ANN model performed slightly better than the benchmark PIM-2 and PRISM-III scores and a traditional logistic regression model trained on the same data set. The small performance gains achieved by this two-layer ANN model may not offer clinically significant improvement; however, further research with other or more sophisticated model designs and better imputation of missing data may be warranted.

IntroductionThe COVID-19 pandemic has exacerbated mental health challenges, particularly depression among college students. Detecting at-risk students early is crucial but remains challenging, particularly in developing countries. Utilizing data-driven predictive models presents a viable solution to address this pressing need.Aims1) To develop and compare machine learning (ML) models for predicting depression in Argentinean students during the pandemic. 2) To assess the performance of classification and regression models using appropriate metrics. 3) To identify key features driving depression prediction.MethodsA longitudinal dataset (N = 1492 college students) captured T1 and T2 measurements during the Argentinean COVID-19 quarantine. ML models, including linear logistic regression classifiers/ridge regression (LogReg/RR), random forest classifiers/regressors, and support vector machines/regressors (SVM/SVR), are employed. Assessed features encompass depression and anxiety scores (at T1), mental disorder/suicidal behavior history, quarantine sub-period information, sex, and age. For classification, models’ performance on test data is evaluated using Area Under the Precision-Recall Curve (AUPRC), Area Under the Receiver Operating Characteristic curve, Balanced Accuracy, F1 score, and Brier loss. For regression, R-squared (R2), Mean Absolute Error, and Mean Squared Error are assessed. Univariate analyses are conducted to assess the predictive strength of each individual feature with respect to the target variable. The performance of multi- vs univariate models is compared using the mean AUPRC score for classifiers and the R2 score for regressors.ResultsThe highest performance is achieved by SVM and LogReg (e.g., AUPRC: 0.76, 95% CI: 0.69, 0.81) and SVR and RR models (e.g., R2 for SVR and RR: 0.56, 95% CI: 0.45, 0.64 and 0.45, 0.63, respectively). Univariate models, particularly LogReg and SVM using depression (AUPRC: 0.72, 95% CI: 0.64, 0.79) or anxiety scores (AUPRC: 0.71, 95% CI: 0.64, 0.78) and RR using depression scores (R2: 0.48, 95% CI: 0.39, 0.57) exhibit performance levels close to those of the multivariate models, which include all features.DiscussionThese findings highlight the relevance of pre-existing depression and anxiety conditions in predicting depression during quarantine, underscoring their comorbidity. ML models, particularly SVM/SVR and LogReg/RR, demonstrate potential in the timely detection of at-risk students. However, further studies are needed before clinical implementation.

Random Forest Modeling of Acute Toxicity in Anal Cancer: Effects of Peritoneal Cavity Contouring Approaches on Model Performance

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.

Coronary artery disease (CAD) results in substantial morbidity and mortality. The purpose of this study was to develop a deep learning model to detect CAD defined using diagnostic codes ("ECG2CAD") and identify people at risk for adverse events using electrocardiograms (ECGs) in a primary care setting. ECG2CAD was trained on 764,670 ECGs representing 137,199 individuals at Massachusetts General Hospital (MGH). Model performance for discrimination of prevalent CAD was measured using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC), and compared against model of age and sex, and Pooled Cohort Equations, in 3 test sets: MGH, Brigham and Women's Hospital (BWH), and UK Biobank. Subgroups were assessed for incident CAD-related events in a BWH primary care cohort. ECG2CAD was evaluated in MGH (N = 18,706 [6,051 cases], age 57 ± 16 years), BWH (N = 88,270 [27,898 cases], age 57 ± 16 years), and UK Biobank (N = 42,147 [1,509 cases], age 65 ± 8 years). ECG2CAD consistently discriminated prevalent CAD (MGH AUROC: 0.782; AUPRC: 0.639; BWH: AUROC: 0.747; AUPRC: 0.588; UK Biobank AUROC: 0.760; AUPRC: 0.155) and incrementally vs models based on age and sex or Pooled Cohort Equations (P < 0.01) in MGH and BWH. In the BWH primary care subset, model performance was consistent across subgroups. Being in the highest quintile of ECG2CAD risk was associated with higher risk for adverse events compared with low-risk group (myocardial infarction HR: 5.59; 95% CI: 4.76-6.56, heart failure 10.49; 95% CI: 7.96-13.84, all-cause mortality 2.68; 95% CI: 2.32-3.10). Artificial intelligence-enabled analysis of the ECG may facilitate identification of individuals with possible undiagnosed CAD and inform downstream testing and preventive measures.

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.

Hospital-onset bacteremia and fungemia (HOB) are common and potentially preventable complications of hospital care. To assess whether nonindividual patient features, which summarize interactions with other patients and health care workers (HCWs), can contribute to predictive and causal machine learning models for HOB. This prognostic study included adult patients admitted to Barnes-Jewish Hospital, an academic hospital in St Louis, Missouri, in 2021. Analyses were developed between October 2023 and August 2024 and in April 2025. Individual patient features were extracted from electronic health records and used to engineer nonpatient features, including interactions with HCWs and direct or indirect (consecutive room occupancy) patient contact. HOB was defined as a positive blood culture after the third day of hospitalization. Patients who were hospitalized for more than 3 days were considered at risk for the outcome. We developed 3 gradient boosting models: 2 predictive (with patient features only and with both patient and nonpatient features to predict the occurrence of HOB) and 1 causal to test the association of nonpatient features and HOB. Predictive performance is reported using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC), and the results of the causal model are reported as difference in average effects. Sensitivity analyses separated intensive care unit-onset and ward-onset HOB and included a methicillin-resistant Staphylococcus aureus-specific model to adjust for colonization pressure. Among the 52 442 patients, 34 855 (66.5%) had admissions longer than 72 hours and were included for analysis; of these, 556 (1.6%) developed HOB. The median age for the included patients was 60 (IQR, 44-70) years, 50.5% were female, and obesity was the most frequent comorbidity (25.0%). Nonpatient features, such as a prior occupant of the same room receiving antipseudomonal beta-lactams and the mean number of HCWs per day for the 7 days preceding HOB, improved the model's performance (AUROC, 0.88 [95% CI, 0.88-0.89]; AUPRC, 0.20 [95% CI, 0.20-0.22]) compared with the patient-only model (AUROC, 0.85 [95% CI, 0.85-0.86]; AUPRC, 0.13 [95% CI, 0.12-0.14]) (P < .001). These 2 features were also associated with a higher likelihood of HOB in the causal gradient boosting model. These findings suggest that nonindividual patient features may contribute to a comprehensive analysis of HOB when integrated with individual patient features in a machine learning model.

The World Health Organization glioma classification highlights genetic profiles, such as isocitrate dehydrogenase (IDH) mutation. This study developed an uncertainty-aware deep learning model to predict IDH mutations in glioma patients, employing conformal prediction (CP) for uncertainty quantification (UQ). The UCSF dataset split into training (70%) and validation (30%) sets, and the UPENN dataset divided into calibration (30%) and external test (70%) sets. We developed various 3D convolutional neural network models and selected the best based on the validation set area under the precision-recall curve (AUPRC). Also, we trained a logistic regression (LR) ensemble classifier on the training set and selected the best ensemble model based on the validation set AUPRC. Finally, we employed CP, with nonconformity thresholds set at the error rate of 0.01. The conformal model was calibrated on the calibration set and was evaluated on the external test set using the AUPRC and the area under the receiver operating characteristic (AUROC). 3D convolutional models achieved a mean AUROC and AUPRC of 0.9004 and 0.8102 on the validation set, and a mean AUROC and AUPRC of 0.7340 and 0.2139 on the external test set, respectively. Using the LR ensemble model, the model achieved an AUROC and AUPRC of 0.79820 and 0.2583 on the external test set, respectively. CP with a 0.01 nonconformity threshold with 0.9917 coverage reached an AUROC and AUPRC of 0.8592 and 0.5217, respectively. Integrating CP for UQ improves the performance of deep learning models for predicting IDH mutation in glioma patients.

Respiratory motion of tumors during radiotherapy can compromise dose delivery and further may impact tumor control efficacy. Developing a predictive model that integrates tumor heterogeneity with respiratory motion may enable the identification of high-risk patients, thereby guiding adaptive radiotherapy strategies and refining follow-up protocols. To propose a novel prognosis prediction method utilizing intra-fractional temporal signatures and validate preliminarily on the prediction of local recurrence (LR) for patients with non-small cell lung cancer (NSCLC). A cohort of 100 NSCLC patients were retrospectively analyzed. The intra-fractional temporal changes of 1321 radiomics features between phase 0% and i were calculated from 4DCT (Delta_RFi, i=10%,.90%) and stacked as accumulative delta-radiomics features (Delta_RFacc). The delta-radiomics signatures (DRSi, i=10%,.90%) and the novel accumulative delta-radiomics signature (ADRS) were developed by applying principal component analysis (PCA) to Delta_RFi and Delta_RFacc. The conventional radiomics signature (RS) was built from 3DCT. The predicting performance using RS, DRSs, and ADRS were compared through univariate Logistic regression. LR risk stratification using ADRS and RS were also evaluated. Based on the mean performance from 7:3 random-split cross-validations with 50 random seeds, the best area under the receiver operating characteristic curve (AUC) and area under the precision-recall curve (AUPRC) were observed in ADRS model (AUC=0.8391, AUPRC=0.8043), which were significantly (all p<0.05) higher than that of RS model (AUC=0.7021, AUPRC=0.5260) and the models based on DRS of any single phase (AUC ranging 0.6470∼0.7693, AUPRC ranging 0.4571∼0.6047). Significantly worse local progression-free survival was associated with higher signature values of both ADRS and RS, with hazard ratios (HR) of 6.74 for ADRS (95% confidence interval (CI): 2.79-16.24, p<0.0001) and 3.28 for RS (95% CI: 1.39-7.75, p=0.0043). The motion-sensitive feature was found contributive to the superior performance of ADRS. The incorporation of temporal information from 4DCT significantly improved the model performance than that of the conventional method based on 3DCT. A novel ADRS approach based on intra-fractional delta-radiomics was proposed and validated on the LR risk prediction for patients with NSCLC. The improved accuracy of risk management can be beneficial to assist the personalized optimization of treatment and follow-up protocols, achieving a balance among cost, efficiency, and outcome.

Central nervous system (CNS) diseases such as glioblastoma (GBM), multiple sclerosis (MS), and Alzheimer's disease (AD) remain challenging due to their complexity and limited treatments. Conventional drug repurposing strategies often rely on bulk RNA sequencing data, which can overlook cellular heterogeneity and mask rare but critical cell populations. Here, we introduce scDrugLink, a computational method that integrates single-cell transcriptomic data with drug targets and perturbation signatures to improve repurposing. For each cell type, scDrugLink constructs a Drug2Cell matrix based on drug targets to estimate promotion/inhibition scores and derives sensitivity/resistance scores by reverse matching signatures and disease-associated genes. These scores are then "linked," yielding robust therapeutic rankings. In our study, we present a systematic evaluation of single-cell drug repurposing methods for CNS diseases. Applied to atlas data for GBM, MS, and AD, scDrugLink surpassed three state-of-the-art methods (ASGARD, DrugReSC, and scDrugPrio), achieving area under the receiver operating characteristic curve (AUC) ranges of 0.6286-0.7242 and area under the precision-recall curve (AUPRC) ranges of 0.3412-0.5484. It also ranked top when comparing AUC and AUPRC at the level of individual cell types. Moreover, applying the "linking" principle to baseline methods boosted their performance, on average improving AUC and AUPRC by 0.0160 and 0.0244, respectively. Despite the advancements, the complexity and heterogeneity of CNS diseases, along with incomplete drug data, indicate that further improvement is necessary. We discuss these challenges and suggest directions for enhancing single-cell drug repurposing in the future.

Purpose: Lung cancer screening (LCS) lowers cancer mortality if individuals adhere to annual scans. Minoritized communities have significantly lower adherence to LCS than non-Hispanic whites. The influence of social determinants of health (SDOH) on these disparate adherence rates are rarely evaluated. The aim of this study is to develop a machine learning (ML) model that integrates individual and neighborhood level SDOH to predict which minoritized individuals are at risk of LCS non-adherence. Methods: A total of 168 participants with high-risk smoking histories who underwent their first low dose computed tomography (LDCT) between 2017 and 2021 were enrolled. Surveys on SDOH and perceived discrimination were completed. Participants’ addresses were joined to neighborhood level SDOH data from various publicly available sources. These indices captured neighborhood socioeconomic status and segregation. Adherence rates were calculated by assessing electronic medical records to determine if participants had a second LDCT within 15 months. Participants were temporally split into training (N=125) and testing (N=43) sets with a cut-off date of June 30, 2021. A gradient boosting classifier, XGBoost, was trained to predict LCS non-adherence, incorporating variables such as sociodemographic status, perceived discrimination, previous screenings, smoking status, and geospatial indices. The SHapley Additive exPlanations (SHAP) method was used to determine the relative importance of each feature in the model’s predictions. Results: The participants represented a diverse group: Asian (26%), Black/African American (32%), and Hispanic/Latino (32%). The LCS non-adherence rates were 66% in the training set and 63% in the testing set. The XGBoost classifier achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.79 and an Area Under the Precision-Recall Curve (AUPRC) of 0.90, outperforming a simple logistic regression model (AUROC=0.63 and AUPRC=0.79). The XGBoost classifier’s performance metrics were Accuracy=0.77, Recall=0.85, Specificity=0.62, Positive Predictive Value=0.79, and Negative Predictive Value=0.71. The SHAP analysis indicated that participants with no prior history of colorectal cancer screening, who lived in areas with lower quality school systems, and with a higher percentage of individuals below the poverty line, were at a higher risk of non-adherence to LCS. Conclusions: LCS reduces lung cancer survival, but LCS adherence is significantly lower in minoritized groups. Our ML model strongly predicted LCS non-adherence based on SDOH and geospatial factors. Specific neighborhood level SDOHs predicted LCS non-adherence more strongly than individual level factors. These findings support interventions that address both individual and neighborhood-level barriers to increase LCS adherence. Citation Format: Kun-Han Lu, Yi Xiao, Aamna Akhtar, Jazma Tapia, Calvin Tribby, Jay Yang, Marta Jankowska, Chi Wah Wong, Loretta Erhunmwunsee. Predicting lung cancer screening adherence among underserved and minoritized communities using social determinants and neighborhood factors [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 4971.

BackgroundMedication adherence is the main determinant of effective management of type 2 diabetes, yet there is no gold standard method available to screen patients with high-risk non-adherence. Developing machine learning models to predict high-risk non-adherence in patients with T2D could optimize management.MethodsThis cross-sectional study was carried out on patients with T2D at the Sichuan Provincial People's Hospital from April 2018 to December 2019 who were examined for HbA1c on the day of the survey. Demographic and clinical characteristics were extracted from the questionnaire and electronic medical records. The sample was randomly divided into a training dataset and a test dataset with a radio of 8:2 after data preprocessing. Four imputing methods, five sampling methods, three screening methods, and 18 machine learning algorithms were used to groom data and develop and validate models. Bootstrapping was performed to generate the validation set for external validation and univariate analysis. Models were compared on the basis of predictive performance metrics. Finally, we validated the sample size on the best model.ResultsThis study included 980 patients with T2D, of whom 184 (18.8%) were defined as medication non-adherence. The results indicated that the model used modified random forest as the imputation method, random under sampler as the sampling method, Boruta as the feature screening method and the ensemble algorithms and had the best performance. The area under the receiver operating characteristic curve (AUC), F1 score, and area under the precision-recall curve (AUPRC) of the best model, among a total of 1,080 trained models, were 0.8369, 0.7912, and 0.9574, respectively. Age, present fasting blood glucose (FBG) values, present HbA1c values, present random blood glucose (RBG) values, and body mass index (BMI) were the most significant contributors associated with risks of medication adherence.ConclusionWe found that machine learning methods could be used to predict the risk of non-adherence in patients with T2D. The proposed model was well performed to identify patients with T2D with non-adherence and could help improve individualized T2D management.

Development, multi-institutional external validation, and algorithmic audit of an artificial intelligence-based Side-specific Extra-Prostatic Extension Risk Assessment tool (SEPERA) for patients undergoing radical prostatectomy: a retrospective cohort study

In an era of expanding sequencing technologies, increased variant identification requires assignment of potential functional impact to prioritize those that may be disease-causing. In this data note, we demonstrate the importance of using a refined human genome reference assembly and more diverse and curated population-based databases in guiding functional annotation of variants identified in inherited retinal disease (IRD) genes. We compared variant characteristics extracted from Genome Aggregation Database (gnomAD) population data extracted for 372 IRD disease genes from versions 3.1.2 (v3) and 4.1.0 (v4), which are aligned to the most recent Genome Reference Consortium Human Build 38 (GRCh38) as well as version 2.1.1 (v2), aligned to the previous GRCh37 build. Transformation of the Variant Effector Prediction (VEP), Combined Annotation Dependent Depletion (CADD) scores, and ClinVar pathogenicity annotations were used to generate receiver-operating characteristic (ROC) curves to calculate area under the curve (AUC) and area under the precision–recall curve (AUPRC). Comparisons of variant prediction by ClinVar designation showed that with improved functional annotation, the AUC climbs to 0.99 and AUPRC is 0.98 in differentiating pathogenic variants from nonpathogenic when using the most recent genome build and population database. More diverse population data allow for identification of rare variants and the incorporation of variant annotation metrics provides greater insight into pathogenicity parameters of IRD variants. This data note provides empirical evidence to adopt the newest genomic builds and databases to better prioritize variants as potentially disease-causing for more complete molecular diagnosis in IRD patients.

Fog computing extends cloud services to the network edge, enabling low-latency processing for Internet of Things (IoT) applications. However, this distributed approach is vulnerable to a wide range of attacks, necessitating advanced intrusion detection systems (IDSs) that operate under resource constraints. This study proposes integrating self-awareness (online learning and concept drift adaptation) into a lightweight RL (reinforcement learning)-based IDS for fog networks and quantitatively comparing it with non-RL static thresholds and bandit-based approaches in real time. Novel self-aware reinforcement learning (RL) models, the Hierarchical Adaptive Thompson Sampling–Reinforcement Learning (HATS-RL) model, and the Federated Hierarchical Adaptive Thompson Sampling–Reinforcement Learning (F-HATS-RL), were proposed for real-time intrusion detection in a fog network. These self-aware RL policies integrated online uncertainty estimation and concept-drift detection to adapt to evolving attacks. The RL models were benchmarked against the static threshold (ST) model and a widely adopted linear bandit (Linear Upper Confidence Bound/LinUCB). A realistic fog network simulator with heterogeneous nodes and streaming traffic, including multi-type attack bursts and gradual concept drift, was established. The models’ detection performance was compared using metrics including latency, energy consumption, detection accuracy, and the area under the precision–recall curve (AUPR) and the area under the receiver operating characteristic curve (AUROC). Notably, the federated self-aware agent (F-HATS-RL) achieved the best AUROC (0.933) and AUPR (0.857), with a latency of 0.27 ms and the lowest energy consumption of 0.0137 mJ, indicating its ability to detect intrusions in fog networks with minimal energy. The findings suggest that self-aware RL agents can detect traffic–dynamic attack methods and adapt accordingly, resulting in more stable long-term performance. By contrast, a static model’s accuracy degrades under drift.