2026-10-09

Timely detection is crucial to improve outcomes in patients with cardiac amyloidosis (CA) by initiation of guideline-directed disease-modifying treatments. Although confirmatory bone scintigraphy is highly accurate for CA detection, identifying at-risk patients for referral remains challenging. This study aimed to develop and validate a machine learning model, Amylo-Detect, using structured multimodal electronic health record (EHR) data to guide referrals for confirmatory scintigraphy and monoclonal protein testing. Consecutive all-comer patients (n = 11,616) referred for bone scintigraphy at the Vienna General Hospital (2010-2023) were retrospectively included. Patients referred before August 2020 formed the development cohort. The remaining patients comprised the internal validation cohort. External validation was performed at the University Hospital Essen (n = 1,521). Amylo-Detect was trained using 50 routinely available parameters to predict CA-suggestive uptake (Perugini grade ≥2) and compared with an existing score and clinical routine. High-grade uptake was present in 388 patients (3.0%). Amylo-Detect demonstrated excellent performance in development (AUC 0.93), independent internal validation (AUC 0.91), and external validation cohort (AUC 0.91), outperforming existing scoring systems and clinical routine. Results were consistent across subgroups, even when crucial predictors were missing. Of the 42/388 (10.8%) patients missed in clinical routine, 12/42 (29%) were additionally detected by Amylo-Detect. The model further conveyed significant prognostic value for mortality and heart failure hospitalization. We present Amylo-Detect, a validated EHR-based tool for CA risk prediction, available as a web app, allowing application and further evaluation. By improving timely detection and referral, Amylo-Detect may help address diagnostic delays, pending prospective evaluation.

Background: Body composition is recognized as a major determinant of health outcomes, but its multidimensional nature makes clinical adoption challenging. We sought to develop and validate a body composition index (BCI) for all-cause mortality risk assessment, integrating variables of six body composition tissues. Methods: We analyzed 28509 consecutive patients undergoing myocardial perfusion imaging with routine low-dose chest CT attenuation correction (CTAC) scans acquired during myocardial perfusion imaging (MPI) at 12 centers across four countries. An artificial intelligence-based BCI was developed in a cohort of 15037 patients' CTACs by integrating the CT-derived metrics of bone, skeletal muscle, and four adipose tissue compartments, coronary artery calcium score, and basic demographic variables (age, sex, BMI). The performance of BCI for mortality prediction was validated in an internal cohort of 6444 patients and an external cohort of 7028 patients by prognosis, calibration, net benefit, and explainability. Model-based simulation of tissue metrics modification was performed to evaluate estimated mortality risk reduction. Findings: During a median of 3.5 (IQR [1.9, 5.1]) years, 4697 (16%) patients died. In the external testing cohort, the BCI demonstrated excellent discrimination for mortality (area under receiver operating characteristic curve 0.78 (95% CI [0.76, 0.79]) and Harrell's concordance index 0.75 [0.73, 0.76]), calibration, and net benefit overall and across pre-specified subgroups stratified by patient characteristics and imaging protocols. Visceral adipose tissue attenuation was the most influential body composition measure, followed by skeletal muscle volume. Simulated improvement in body composition was associated with significant mortality risk reduction. Interpretation: An index combining six body composition measures obtained opportunistically from routine chest CT provides robust mortality risk stratification. By converting complex body composition information into a single interpretable score, the BCI can facilitate clinical implementation of opportunistic CT biomarkers and guide individualized preventive strategies. Funding: National Institute of Health. Research In Context: Evidence before this study: We searched PubMed and Google Scholar on July 7, 2026 for English language studies using the terms ("mortality prediction" OR "death prediction") AND ("integrated body composition analysis" OR "body composition metrics integration") AND ("predictive model" OR "model development"), and identified 41 publications describing predictive models incorporating body composition metrics.Recent advances in artificial intelligence (AI) have enabled increasingly granular quantification of body composition from CT, with multiple tissues and tissue-specific metrics showing independent prognostic value for all-cause mortality (ACM) prediction. However, reported associations of body composition metrics with mortality risk were highly inconsistent or even contradictory across studies. This heterogeneity likely reflected the differences in study design or the inherent complexity of body composition analysis, such as the interaction among variables derived from multiple tissues. Thus, it is difficult to interpret body composition results reliably in clinical practice. We hypothesized that integrating multiple body composition metrics could achieve comprehensive and reliable prediction of mortality across patient populations.Added value of this study: In this large international, multicenter longitudinal cohort study of 28509 consecutive patients with suspected or known coronary artery disease (CAD) undergoing myocardial perfusion imaging across 12 sites in four countries, we developed and externally validated an artificial intelligence-based body composition index (BCI) for all-cause mortality prediction. The BCI integrated chest CT-derived metrics of six body composition tissue compartments - bone, skeletal muscle (SM), subcutaneous adipose tissue (SAT), intramuscular adipose tissue (IMAT), visceral adipose tissue (VAT), and epicardial adipose tissue (EAT), together with demographics (age, sex, BMI) and CT-derived coronary artery calcium (CAC) score, all derived automatically from routine low-dose attenuation correction CT scans without additional imaging or laboratory tests.In an external validation cohort of 7028 patients, the BCI achieved robust discrimination for mortality prediction, substantially outperforming a baseline model using demographics and calcium score alone, and exceeded the discriminative performance of prior body composition models. Performance was consistent across subgroups defined by sex, age, body mass index, cardiometabolic risk factors, imaging modality, and acquisition protocol, supporting generalizability across clinical settings. The BCI demonstrated excellent calibration and net clinical benefit. Explainability analysis identified visceral adipose tissue attenuation and skeletal muscle volume as the dominant contributors to mortality risk, offering tissue-specific therapeutic targets that complement generic weight-based metrics. Model-based simulations suggested that improvements in these specific compartments could reduce estimated mortality risk more effectively than equivalent changes in BMI alone, pointing toward precision therapeutic targets beyond weight reduction.Our study improved over previous literatures on body composition-based predictive models in one or more of the following ways: (1) we considered six types of body composition tissues while existing studies considered only three or less tissues, thus significantly higher granularity and improved ability to capture refined tissue interactions; (2) our model was rigorously validated in external sites, showing cross-site generalizability while most previous work lacked external validation; (3) our model achieved significant prediction performance improvement over previous body composition predictive models, offering better clinical outcomes prediction and facilitating the suitability for widespread clinical adoption; (4) our study was conducted with a large multi-center cohort across multiple countries, achieving statistical power and demonstrating reliability towards patients characteristics and imaging acquisition protocols while previous studies involved small single-center cohorts; (5) our model required only CT scans and basic demographics which were widely available in clinical routine while the existing models required dedicated laboratory tests or specialized imaging modalities, enhancing the applicability in routine clinical practice.Implications of all the available evidence: BCI offers an explainable and clinically deployable tool for improved prediction of mortality risk from existing chest CT scans. Its use could inform updates to existing risk models, guide precision preventive strategies, and support development of therapeutics targeting specific body composition phenotypes.

BackgroundPostoperative delirium is a serious complication in elderly ICU patients after gastric surgery, increasing mortality and healthcare burden. Early identification of high-risk patients may support timely prevention and optimise perioperative care.ObjectiveTo develop and externally validate an interpretable machine learning model for predicting postoperative delirium in elderly ICU patients undergoing gastric surgery.MethodsThis retrospective multicohort study used three perioperative and critical care databases. MIMIC-IV supported model development and internal validation, while eICU and MOVER supported external validation. Feature selection used LASSO regression, followed by comparison of six machine learning algorithms. Logistic regression was selected using the MIMIC-IV internal validation cohort based on discrimination, recall, and clinical interpretability; the locked model was then evaluated in eICU and MOVER. SHAP analysis quantified predictor contributions and enhanced model transparency.ResultsThe final population included four cohorts: MIMIC-IV training (N = 4316), MIMIC-IV internal validation (N = 1851), eICU external testing (N = 1745), and MOVER external testing (N = 1152). The internally selected logistic regression model achieved AUCs of 0.819 in eICU and 0.827 in MOVER, with recalls of 0.827 and 0.936, respectively. SHAP analysis identified chloride, sodium, bicarbonate, and anion gap as dominant predictors across cohorts.ConclusionThis study presents an interpretable postoperative delirium risk-prediction model in elderly ICU patients following gastric surgery. Although the model achieved high recall, marked class imbalance and low outcome prevalence were associated with low precision and a substantial false-positive burden. Prospective validation, threshold optimisation, and assessment of clinical utility are required before implementation.

Background: Low- and middle-income countries (LMICs), particularly Sub-Saharan Africa, experience a high burden of surgical disease and a disproportionately high postoperative morbidity and mortality. Although machine learning based pediatric surgical risk calculators (SRCs) have been developed to support clinical decision making, most are derived from high-income settings and may not generalize well to LMIC populations. We previously developed a pediatric surgical risk calculator using multinational LMIC data; however, its performance in independent African cohorts remains unknown. Methods: We conducted an external validation study of a previously developed pediatric surgical risk calculator using prospective data from children undergoing surgery at four tertiary referral hospitals in Ethiopia and Nigeria between July 2024 and February 2025. Model performance was assessed by evaluating discrimination (area under the receiver operating characteristic curve) and calibration (slope and intercept). Predicted risk was stratified into low, intermediate, and high-risk categories to enhance clinical applicability. Model recalibration was performed to optimize performance in the external cohort. Results: A total of 1,670 pediatric surgical patients were included from the four hospitals. In-hospital mortality was 6.8%. Mortality varied significantly by site, age, weight, sex, prematurity, preoperative sepsis, transfusion, supplemental oxygen use, antibiotic use, ASA class, type of surgery (emergent vs elective), need of ICU postoperatively and reoperation. The risk prediction model had good discrimination overall with an area under the curve (AUC) of 0.84, outperforming ASA classification alone (AUC = 0.77). Discrimination remained high for general pediatric surgery cases (AUC=0.83) and was highest for emergency cases (AUC = 0.89), but was low in neonates (AUC = 0.64). The model demonstrated systematic miscalibration, which improved following logistic calibration. Patients were stratified into meaningful risk categories, which reliably distinguishes between low-and-hgih risk patients (LR=0.1 and LR=5.4 respectively). However, 66.5% fall in the intermediate risk group which is not useful in decision making (LR=1.0). Conclusion: The pediatric surgical risk calculator demonstrated good performance in the African context, with improved clinical utility following recalibration and risk stratification. Reduced performance in neonates suggests the need to incorporate additional risk factors and potentially separate models for this subgroup. Future efforts should focus on supporting the use of the SRC in risk communication, clinical decision making, and resource prioritization.

Purpose: Accurate preoperative assessment of lymphovascular space invasion (LVSI) in cervical cancer (CC) remains challenging despite its importance for treatment planning and risk stratification. This study aimed to develop and evaluate a hybrid model integrating peritumoral ultrasound radiomics features, deep learning features, and clinical data for preoperative prediction of LVSI in CC. Additionally, we assessed the model's prognostic value for disease-free survival (DFS) and explored LVSI-associated molecular pathways using data from The Cancer Genome Atlas (TCGA). Methods: In a retrospective cohort (n = 241) and a prospective cohort (n = 51), radiomics features were extracted from tumor and peritumoral regions of interest on ultrasound images. Least absolute shrinkage and selection operator regression and machine learning classifiers generated a radiomics score (Rad-score). Simultaneously, transfer learning with pre-trained convolutional neural networks constructed a deep learning score (DL-score). A hybrid model integrating clinical factors, Rad-score, and DL-score was developed. Model performance was evaluated using the area under the curve (AUC). Furthermore, prognostic value (Kaplan-Meier analysis), model interpretability using Shapley Additive Explanations (SHAP), and biological pathways were assessed. Results: Overall, 292 women were included: 188 in the training set, 53 in the internal validation set, and 51 in the prospective testing set. The hybrid model achieved the highest performance (AUC: 0.79, 95%CI: 0.66-0.92; sensitivity: 0.85, 95%CI: 0.71-0.98) in the prospective testing set, outperforming the single-modality models. Exploratory survival analysis showed shorter DFS in the high-risk group (P < 0.05). SHAP analysis identified the DL-score as the primary predictor. Biologically, the Wnt/β-catenin pathway was significantly enriched in LVSI-positive patients. Conclusions: The hybrid model, evaluated in a prospective cohort, showed potential for preoperative LVSI prediction and DFS risk stratification in CC. TCGA analysis showed Wnt/β-catenin enrichment in LVSI-positive tumors, providing complementary biological context. Further validation in larger independent multicenter cohorts is needed before clinical application.

Background: Metabolic dysfunction-associated steatotic liver disease (MASLD) is common among obese individuals undergoing bariatric surgery, but liver biopsy is invasive and existing noninvasive tools generally assess single histologic features. We aimed to develop and externally validate a simple framework for predicting multiple liver histologic injuries. Methods: We included 909 obese individuals undergoing sleeve gastrectomy with intraoperative liver biopsy in the training cohort and 186 temporally separated patients in the internal test cohort. Six separate outcome-specific binary classification models were developed, one for each histologic outcome: steatosis, hepatocyte ballooning, significant fibrosis, metabolic dysfunction-associated steatohepatitis (MASH), NAS ≥5, and at-risk MASH. Multiple machine-learning algorithms were evaluated, and the best-performing algorithm was selected for each outcome. A simplified three-variable TabPFN framework using alanine aminotransferase, homeostasis model assessment of insulin resistance, and peripheral monocyte ratio was developed, with one binary model for each outcome. External validation included two biopsy-confirmed cohorts (n = 103 and 217) and the U.S. NHANES cohort (n = 4,299) using elastography-based assessments. Results: The models demonstrated good discrimination for multiple histologic liver injuries in the test cohort, with ROC-AUCs ranging from 0.73 to 0.86 across outcomes. The simplified three-variable framework showed consistent external performance, achieving ROC-AUCs of 0.84 for ballooning, 0.86 for steatosis, 0.79 for significant fibrosis, and 0.75 for MASH. The framework demonstrated generalizability in the NHANES population. Conclusions: This externally validated framework may support noninvasive identification of multiple liver histologic injuries and perioperative liver risk stratification in individuals undergoing bariatric surgery, thereby improving the clinical accessibility of liver pathology assessment.

Recovery following anterior cruciate ligament reconstruction (ACLR) is multifaceted, encompassing strength, functional, and patient-reported outcome measures (PROMs). Determining which factors are most important for characterizing recovery and informing return-to-activity (RTA) decisions remains a clinical challenge. This study aimed to develop and externally validate a supervised machine learning model to classify individuals post-ACLR from healthy controls using comprehensive lower-extremity musculoskeletal assessments. Data from 863 participants (659 post-ACLR, 204 controls) at Site 1 were used to train and internally test Random Forest, Gradient Boosting Machine, and Extreme Gradient Boosting (XGBoost) models, with an 80/20 training-testing split and inner fivefold cross-validation for hyperparameter optimization. Model performance was evaluated using accuracy and area under the receiver operating characteristic curve (AUC), and external validation was performed on an independent Site 2 cohort (N = 174; 133 post-ACLR, 41 controls). The XGBoost model was selected for further analysis (internal test AUC = 0.98, Accuracy = 0.95), and maintained good discrimination on external validation (AUC = 0.83, Accuracy = 0.82). The most informative features included the Knee Osteoarthritis Outcome Score quality-of-life subscale, limb symmetry index for peak knee extension torque at 90°/s, and the International Knee Documentation Committee score, which were consistent across age and sex subgroups. These findings demonstrate that machine learning can effectively integrate strength, function, and PROM data to identify key indicators of ACLR status and support development of streamlined, data-driven protocols for post-ACLR recovery monitoring and RTA decision-making. STATEMENT OF CLINICAL SIGNIFICANCE: The Knee Osteoarthritis Outcome Score quality-of-life subscale, limb symmetry index for peak knee extension torque at 90°/s, and the International Knee Documentation Committee score most effectively distinguish individuals following ACLR from healthy controls. Clinicians may focus on these tests to determine ACLR recovery status as a feasible streamlined assessment battery or when results from a comprehensive testing battery are conflicting or inconclusive.