A new simplified minimum data set better predicts outcomes for individuals admitted to intermediate care functional recovery units in catalonia.

Noninvasive intracranial hypertension detection using machine-learning of cerebral blood flow velocity waveforms

Fever of unknown origin (FUO) remains a complex diagnostic challenge, with infection, malignancy, and inflammatory diseases as the leading causes. A substantial proportion of FUO patients present with lymphadenopathy, among which lymphoma and benign lymph node disorders are the most critical differential diagnoses due to their markedly different management and prognosis. Although 18F-FDG PET/CT plays an indispensable role in lymphoma detection, significant metabolic overlap between malignant and benign lymphadenopathy limits its standalone diagnostic accuracy in FUO. Therefore, integrating PET/CT metabolic parameters with clinical features and laboratory markers may improve the differentiation between lymphoma and benign lymphadenopathy, facilitating timely and accurate diagnosis in FUO patients. This study retrospectively analyzed patients with FUO who underwent PET/CT. Lymph nodes were included if their short-axis diameter was ≥ 1cm on axial CT or if metabolic activity exceeded that of the mediastinal blood pool. Volumes of interest were manually delineated in LIFEx using a 40% SUVmax threshold. Among 203 patients (114 with lymphoma and 89 with benign lymphadenopathy), logistic regression analyses identified independent predictors across three domains: hyperpyrexia, joint pain, and rash among clinical factors; albumin (ALB), procalcitonin (PCT), and serum amyloid A (SAA) among laboratory indicators; and SUVmean, total metabolic tumor volume (TMTV), and total lesion glycolysis (TLG) among PET parameters. A combined model incorporating rash, PCT, SAA, SUVmean, and TMTV achieved superior diagnostic performance, with an area under the ROC curve (AUC) of 0.965, significantly outperforming models based solely on clinical (AUC = 0.790), laboratory (AUC = 0.866), or PET (AUC = 0.845) variables (all P < 0.001). A nomogram was subsequently developed for individualized risk prediction. Rash, PCT, SAA, SUVmean, and TMTV were independent predictors distinguishing lymphoma from benign lymphadenopathy in FUO. Integrating clinical, laboratory, and PET parameters markedly improved diagnostic accuracy.

High-grade serous ovarian cancer (HGSOC) remains a lethal gynecologic malignancy with limited tools for accurately predicting PARP inhibitor (PARPi) efficacy. While promoter methylation of homologous recombination repair (HRR) genes is a potential biomarker, its clinical utility remains underexplored. This study aimed to systematically evaluate HRR gene promoter methylation profiles to predict PARPi sensitivity in HGSOC patients. We conducted a retrospective analysis of 96 HGSOC patients receiving PARPi maintenance therapy (olaparib or niraparib). The promoter methylation status of 12 HRR genes was quantitatively assessed via pyrosequencing. Patients were stratified into PARPi-sensitive and -insensitive groups based on OReO/ENGOT-ov38 criteria. Differential methylation sites were identified, lasso and multivariate logistic regression identified key methylation sites to construct a nomogram and a PARPi Resistance Score (PRS). PFS was analyzed by Kaplan‑Meier and Cox regression. The PARPi-insensitive group exhibited significantly higher median methylation levels across HRR gene promoters compared to the sensitive group. Through rigorous screening, 12 CpG sites were consistently associated with PARPi response across subgroups. Through screening and LASSO regression, five CpG sites-ATM_Pos1, BRCA1(2S)_Pos2, BRCA2(1S)_Pos1, BRCA2(1S)_Pos7, and RAD51D_Pos6-were identified as independent predictors of PARPi sensitivity. The logistic model based on these five sites showed good discrimination, with area under the ROC curve (AUC) of 0.847 in the training set and 0.903 in the validation set; calibration and decision curve analysis supported clinical utility. Using the optimal PRS cutoff of -1.023, High-PRS patients had significantly shorter median PFS (14months vs. not reached; log‑rank P < 0.0001) and lower 12‑month PFS rate (52% vs. 97.2%). Multivariate Cox analysis confirmed PRS as the only independent prognostic factor for PFS (HR 4.85, 95% CI 2.31-10.20, P < 0.001). HRR gene promoter methylation, particularly at specific BRCA1, BRCA2 and RAD51D CpG sites, robustly predicts PARPi sensitivity and independently stratifies PFS in HGSOC patients. The nomogram and PRS provide clinically actionable tools for individualized PARPi therapy.

This study presents a transformer-based Siamese framework for iris verification that integrates image enhancement, attention-guided segmentation, and global feature learning within a unified pipeline. A deep denoising autoencoder is employed to improve image quality, followed by an attention-guided U-Net for precise iris localization. Global feature representations are extracted using a Vision Transformer, enabling the modeling of long-range dependencies across iris textures, and are optimized through contrastive learning to enhance discriminability. The proposed framework is evaluated on the CASIA-IrisV3 dataset, including the Interval, Lamp, and Twins subsets, under a subject-disjoint verification protocol. Experimental results achieve an Equal Error Rate (EER) of 2.34% and an Area Under the ROC Curve (AUC) of 0.987, with a True Acceptance Rate of approximately 95% at a False Acceptance Rate of 10⁻³. The method attains a verification accuracy of 97.82% while maintaining efficient inference with an average latency of 12.6 ms per comparison. These results indicate consistent performance across varying acquisition conditions within the near-infrared domain, demonstrating the effectiveness of integrating attention-guided segmentation and transformer-based representations for iris verification.

Venous thromboembolism (VTE) is a growing public health threat whose prevalence imposes a significant burden on patients and the economy. It mostly affects the lower limbs and is more prevalent in patients who have fractures or who are bedridden for long periods of time. Deep vein thrombosis (DVT) is more common and advances more quickly in cases of traumatizing lower extremity fractures. This study aims to develop a nomogram to predict the risk of deep vein thrombosis (DVT) progression following lower limb fracture surgery, providing a theoretical foundation for preoperative prevention. This retrospective study analyzed 500 patients who underwent lower limb fracture surgery at the Third Hospital of Hebei Medical University. The mean age was 54 years, and 65.8% were male. Thrombotic progression occurred in 30.2% of patients. Data collected included demographics, comorbidities, surgical details, and laboratory parameters. Independent risk factors for postoperative DVT progression were identified using logistic regression. A nomogram prediction model was developed and evaluated using the area under the ROC curve (AUC), calibration curves, and decision curve analysis (DCA). Results showed 5 significant (P < .05) independent risk factors for thrombosis progression: age (45–59 years), blood transfusion, short time from fracture to surgery, elevated d-dimer levels, and preoperative thrombosis location (intermuscular vein thrombosis). Using these 5 independent criteria, a risk prediction model for DVT progression in patients following lower extremity fracture surgery exhibited good performance with an AUC value of 0.691 (95% CI: 0.642–0.740). Internal validation of the model revealed that the calibration curves were near the ideal curves. The developed and validated prediction model demonstrated good accuracy in identifying high-risk patients and provided key insights into DVT pathogenesis. This reliable tool informs clinical strategies for targeted interventions, ultimately improving patient outcomes.

Breast lesion segmentation and classification in ultrasound (US) images are two essential tasks for computer-aided diagnosis of breast cancer. However, these tasks are still challenging, mainly due to the high variability of lesions and the poor image quality of US. Numerous deep learning methods have been proposed to assist physicians in performing breast lesion segmentation and classification. Considering that these tasks are related to common features, learning them jointly through multi-task learning (MTL) represents a viable approach to improve the performance of each individual task. In this paper, we present and compare multiple MTL network configurations for the simultaneous segmentation and classification of breast lesions in ultrasound images. Building on two state-of-the-art architectures, namely SegResNet for segmentation and EfficientNet for classification, we designed and evaluated several combined configurations of these models arranged in different MTL schemes. These configurations explore various levels of feature sharing and integration between the tasks, aiming to identify the most effective architectural arrangement for joint lesion segmentation and malignancy classification in breast ultrasound. Moreover, these configurations are also compared with state-of-the-art MTL configurations. Experimental results based on a dataset compiled from two different centers, comprising 810 2D breast US images, indicate that combining SegResNet and EfficientNet in a MTL configuration achieved accurate performance in both segmentation and classification. Among the tested configurations, the best-performing model combined a segmentation network with an EfficientNet classification network that utilized shared encoded features from the SegResNet, achieving a Dice coefficient of 81.19% for lesion segmentation and an area under the ROC curve (AUC) of 97.27% for lesion classification. Overall, the proposed networks demonstrated their potential in enhancing computer-aided diagnosis of breast cancer.

Flash floods represent one of the deadliest weather-related hazards globally, yet their prediction remains fundamentally challenged by extreme class imbalance in observational data. This study addresses a critical methodological gap: traditional evaluation metrics, both overall accuracy and Area Under the ROC Curve (AUC), are systematically misleading for rare event prediction. We demonstrate empirically how models achieving 93% accuracy and AUC exceeding 0.98 can simultaneously fail to detect 65% of flood events. Moving beyond conventional approaches, we introduce distribution theory-informed feature generation by integrating Extreme Value Theory through Weibull distribution analysis. We derive 24 features from rigorous statistical characterization of precipitation extremes spanning 16 years (2010-2026) of ERA5-Land reanalysis over Nova Scotia, Canada. Evaluating seven model configurations using Environment and Climate Change Canada operational warning thresholds, we find that adding just six Weibull-derived features to a Random Forest baseline nearly doubles flood detection, with recall increasing from 0.35 to 0.65 and F1-score from 0.48 to 0.74, while maintaining 87% precision. This controlled comparison provides the clearest evidence for the value of distribution-informed features. Across architectures, Support Vector Machines with selected features achieve 93.4% balanced accuracy with perfect recall, while Artificial Neural Networks achieve a balanced operational profile (75% recall, 65% precision). SHAP analysis reveals that physically meaningful interaction features, particularly the intensity-duration product and rain-on-saturated-soil, dominate predictions, with raw precipitation ranking only sixth, confirming that models learn genuine multivariate susceptibility structure rather than recovering classification thresholds. These findings provide essential guidance for practitioners: comprehensive reporting of balanced accuracy, precision, and recall is mandatory for imbalanced datasets where traditional metrics mask operational failure.

ABSTRACTAimTo investigate the risk factors for postoperative hypoproteinemia in patients undergoing microsurgical reconstructive surgery and to develop a prediction model to support nurses in identifying patients at high risk.DesignA prospective observational study.MethodsA convenience sample of 192 patients with limb‐destructive injury admitted to the Department of Orthopaedics at a tertiary hospital between September 2021 and March 2023 was enrolled. Patients were classified into hypoproteinemia (serum albumin < 35 g/L) and non‐hypoproteinemia (serum albumin ≥ 35 g/L) groups based on serum albumin levels measured within the first 72 h postoperatively. Multivariable logistic regression was used to develop a predictive model. Receiver operating characteristic curve analysis was performed to assess model discrimination performance.ResultsThe prevalence of postoperative hypoproteinemia in this cohort was 72.4%. Statistically significant between‐group differences were observed for body mass index, smoking status, wound size, operating time, intraoperative blood loss, amount of fluid rehydration on the day after surgery, preoperative total protein level, preoperative albumin level, preoperative red blood cell count, preoperative haematocrit (Hct), preoperative haemoglobin (Hb) level, and the Mangled Extremity Severity Score (MESS). Multivariable logistic regression analysis showed that higher preoperative albumin level (odds ratio [OR] = 0.837 per 1 g/L increase) and higher preoperative Hb level (OR = 0.958 per 1 g/L increase) were independently associated with lower odds of postoperative hypoproteinemia, whereas longer operating time (OR = 1.967 per 1h increase) was associated with increased odds. The area under the ROC curve (AUC) for the combined model was 0.915.Patient or Public ContributionNo patient or public contribution.

The comorbidity rate of depression and anxiety disorders is as high as 50%, with patients suffering from both conditions facing an elevated risk of suicide. Given the challenges in clinical diagnosis, there is an urgent need for objective diagnostic biomarkers to accurately identify patients with comorbid depression and anxiety disorders. The case-control study enrolled 77 participants, including healthy controls (n = 38) and patients with comorbid depression and anxiety disorders (n = 39). Serum levels of short peptides were assessed using liquid chromatography-tandem mass spectrometry (LC-MS/MS). Binary logistic regression and receiver operating characteristic analysis (ROC) were employed to identify potential biomarkers, while decision curve analysis (DCA) was performed to evaluate the clinical utility of the candidate cyclic pentapeptide. A total of 25 differentially expressed short peptides were identified, among which 19 were significantly downregulated, and 6 were upregulated. Logistic regression analysis revealed that reduced levels of cyclo (D-trp-D-asp-pro-D-val-leu) were associated with an increased risk of comorbid depression and anxiety disorders. The area under the ROC curve (AUC) for this cyclic pen-tapeptide in diagnosing comorbid depression and anxiety disorders was 0.941 (95% CI, 0.884-0.997; P < 0.0001). DCA demonstrated that the cyclic pentapeptide provides meaningful net benefits when the threshold probability exceeded 0.11. Notably, this cyclic pentapeptide is an endothelin A receptor antagonist, which is implicated in vasoconstriction, hypertension, and anxiety, and elevated blood pressure was indeed observed in the patient cohort. The identified cyclic pentapeptide serves as a potential diagnostic tool, linking the pathophysiology of comorbid depression and anxiety disorders to the endothelin system. This association offers a novel integrative perspective on the co-occurring psychiatric and cardiovascular manifestations observed in these patients. In conclusion, this study identifies a specific cyclic pentapeptide as a promising serum biomarker that may facilitate the objective diagnosis of patients with comorbid depression and anxiety disorders. Its biological role as an endothelin A receptor antagonist not only provides a potential tool for clinical differentiation but also suggests a novel pathophysiological link to cardiovascular comorbidities, thereby offering new directions for elucidating the underlying mechanisms of these conditions.

Risk stratification of MASLD-related advanced fibrosis using a novel point-of-care device: the Hepatoscope.

Forest cover restoration is urgently needed in a semi-arid district of West Bengal, where land degradation endangers environmental stability and community welfare. The present study introduces and validates a robust, data-driven framework using machine learning to isolate optimal sites for afforestation, aiming to enhance climate adaptability and create sustainable, forest-centric livelihood opportunities. The methodology is structured as a sequential, hybrid workflow. Initially, an unsupervised K-Means clustering algorithm was applied to a suite of eleven environmental variables derived from SRTM, Landsat, and national geospatial databases to perform an exploratory delineation of potential zones. This was followed by a meticulous training data generation were manually digitized through high-resolution visual validation on Google Earth Pro. This dataset then served as the basis for training two supervised algorithms: RF and XGBoost. A rigorous comparative evaluation confirmed the superior predictive power of the Random Forest model, which achieved an overall accuracy of 89.1% and Area Under the ROC Curve (AUC) of 0.9508. An interpretability analysis using SHAP further revealed that slope, soil moisture, and elevation were the most critical determinants of suitable area. The primary outcome is spatially explicit suitability map with 20.9% area of the district as potentially suitable for afforestation that serves as a decision-support tool, enabling policymakers and community stakeholders to implement strategic and effective afforestation programs in the study area.

ABSTRACT Objective This study aimed to investigate the correlations of serum uric acid (SUA) and glycated hemoglobin (HbA1c) levels with mild cognitive impairment (MCI) in patients with comorbid type 2 diabetes mellitus (T2DM) and hypertension, and to further identify potential risk factors associated with MCI. Methods In this retrospective study, 126 older patients with T2DM and hypertension were divided into a normal cognition group (n = 74) and an MCI group (n = 52) based on Mini-Mental State Examination (MMSE) scores. The correlations of SUA and HbA1c with MMSE scores, their predictive value for MCI, their relationships with clinical variables, and independent risk factors for MCI were evaluated using Spearman correlation, receiver operating characteristic curve, and logistic regression analyses. Results In the MCI group, SUA and HbA1c levels were significantly negatively correlated with MMSE scores (r = -0.419 and −0.510). For predicting MCI, the area under the ROC curve (AUC) was 0.742 for SUA (sensitivity: 78.85%, specificity: 62.16%, cut-off value: 319.31), 0.781 for HbA1c (sensitivity: 48.08%, specificity: 94.59%, cut-off value: 8.94), and 0.851 for their combination (sensitivity: 73.08%, specificity: 85.14%, cut-off value: 0.42). Elevated SUA and HbA1c levels and older age were independent risk factors for MCI, whereas higher IBIL and HDL-C levels and reading habits were protective factors. Conclusions SUA and HbA1c levels were elevated in older patients with T2DM and hypertension who developed MCI. These markers were identified as independent risk factors for MCI in patients with T2DM-hypertension, aiding in the prediction of MCI occurrence in these patients.

Background Emergency ventral hernia repair remains a challenging procedure due to patient instability, contaminated surgical fields, and heterogeneity in hernia types and operative techniques. Predicting postoperative complications in this setting is difficult using traditional statistical methods. Machine learning (ML) may offer improved predictive accuracy by recognizing nonlinear patterns among multiple perioperative factors. Methods A retrospective multicenter analysis was performed using data from the ACTIVE (Acute Treatment for Incisional Ventral Hernias) study, including 557 adult patients undergoing emergent ventral hernia repair between 2018 and 2021 in 31 Italian surgical centers. Demographic, preoperative, intraoperative, and postoperative variables were analyzed. Three ML algorithms—Decision Tree, Random Forest, and Deep Learning Neural Network—were trained and validated using five-fold cross-validation after class balancing with SMOTE. Model performance was compared with traditional logistic regression using accuracy, area under the ROC curve (AUC), and F1 score. Results Postoperative complications occurred in 181 patients (32.5%), while major complications (Clavien–Dindo ≥ II) occurred in 10%. Random Forest achieved the best performance (AUC 0.95, accuracy 0.88, F1 score 0.86), outperforming logistic regression (AUC 0.82, accuracy 0.78). The most influential predictors were operative duration, ASA score, and sepsis for overall complications, while bowel obstruction and BMI were key factors for major complications. Surgical approach (open vs. laparoscopic) did not independently correlate with adverse outcomes, highlighting the complexity of patient- and case-specific interactions. Conclusions Machine learning models can accurately predict postoperative complications following emergent ventral hernia repair, surpassing traditional regression methods. These findings suggest that ML-based decision tools could support risk stratification and optimize surgical planning in high-risk emergency settings. Prospective validation is warranted to integrate AI-assisted prediction into perioperative clinical workflows.

Postoperative delirium is a frequent complication after cardiac surgery. The predictive value of initial lactate at ICU admission for cardiac surgery-associated postoperative delirium (CS-POD) remains unclear. We conducted a retrospective cohort study using the eICU Collaborative Research Database, including adult patients admitted to the ICU after cardiac surgery. Multivariable logistic regression, restricted cubic spline analysis, ROC analysis, and mediation analysis were performed to assess the association between initial lactate and CS-POD and to explore whether disease severity scores (SOFA, APACHE IV) or clinical interventions statistically explained this relationship. Among 358 patients, 104 (29.1%) developed CS-POD. Higher initial lactate at ICU admission was independently associated with increased risk of CS-POD after full adjustment (OR per 1mmol/L increase: 1.37, 95% CI 1.11-1.69; P = 0.003). Restricted cubic spline analysis demonstrated a linear relationship between initial lactate and CS-POD risk (P for nonlinearity = 0.132). Adding initial lactate to the baseline prediction model improved discriminative performance, with the area under the ROC curve (AUC) increasing from 0.661 (95% CI, 0.598-0.723) to 0.717 (95% CI, 0.660-0.775) (DeLong test P = 0.012). The optimal lactate cutoff for predicting CS-POD was 2.385mmol/L. Mediation analysis indicated that part of the association between initial lactate and CS-POD was statistically explained by SOFA (proportion statistically explained: 17%, P = 0.002) and APACHE IV (proportion statistically explained: 7.7%, P = 0.034), whereas clinical interventions (IABP, RRT, opioid use) did not show significant mediation. Higher initial lactate levels were also associated with increased risks of acute kidney injury (OR 1.27, 95% CI 1.00-1.61) and in-hospital mortality (OR 2.88, 95% CI 1.52-5.46). Higher initial lactate at ICU admission is associated with an increased risk of CS-POD and adverse postoperative outcomes. However, its predictive value is modest and should be interpreted cautiously as part of a broader clinical assessment rather than as an independent decision-making tool.

BackgroundNurses with neck–shoulder pain are at high risk of developing limited shoulder range of motion (ROM), but early prediction tools are lacking. This study aimed to identify risk factors using clinical data and musculoskeletal ultrasound (MSKUS), and to develop and validate a prediction model.MethodsA total of 210 nurses with neck–shoulder pain (100 with limited ROM, 110 with normal ROM) from a tertiary hospital were enrolled. Univariate and multivariate logistic regression analyses were performed to establish a prediction model, which was externally validated in 63 additional nurses.ResultsMultivariate analysis identified age, average sleep duration on workdays (ASDW), daily overhead arm duration (DOAD), glenohumeral joint effusion (GLE), subacromial-subdeltoid bursa effusion (SASDBE), and long head of biceps tendon sheath effusion (LHBTE) as independent risk factors. The prediction model demonstrated excellent discrimination with an area under the ROC curve (AUC) of 0.956, sensitivity of 0.910, and specificity of 0.864. External validation yielded an AUC of 0.938, sensitivity of 0.833, and specificity of 0.821.ConclusionThe model incorporating age, ASDW, DOAD, GLE, SASDBE, and LHBTE shows excellent predictive performance for limited shoulder ROM in nursing staff with neck–shoulder pain.

Introduction: Obesity and overweight are important risk factors for a wide range of diseases, including high blood pressure. In addition, anthropometric indicators are considered independent predictors of blood pressure. The purpose of this study was to investigate the relationship between blood pressure and anthropometric indices and to determine the best determinant of hypertension in the Yazd Shahdieh cohort population. Methods: The present study was a descriptive study conducted during the first phase of Shahdieh Cohort Study on, involving 10194 residents of Shahdieh, Ashkazar and Zarch Cities. eligible individuals were enrolled, and the data were analyzed using STATA 16 through Chi-square, T-test, ANOVA, logistic regression, and ROC Curve analyses. Results: Among the participants, 36.7% of adults were found to have high blood pressure. Adjusted logistic regression results showed that the variables of gender, BMI, age, education level, triglyceride levels, diabetes, and family history of hypertension in first-degree relatives as significant predictors of hypertension. Of all anthropometric indices, the waist-to-hip ratio demonstrated the highest diagnostic value and accuracy for determining blood pressure, with an area under the ROC curve (AUC) of 0.67. Conclusion: Considering the observed relationship between anthropometric indicators and elevated high blood pressure, it is evident that promoting education on obesity prevention, along with lifestyle modification, healthy nutrition, and regular physical activity, is essential to reduce the risk and prevalence of hypertension.

Abnormal bone turnover in chronic kidney disease (CKD) increases fracture risk and influences treatment response, making accurate assessment essential. Although bone histomorphometry is the diagnostic gold standard, its invasiveness limits routine use. This study evaluated the diagnostic accuracy of noninvasive bone turnover markers-tartrate-resistant acid phosphatase 5b (TRAP5b), bone-specific alkaline phosphatase (BALP), and procollagen type I N-terminal propeptide (PINP)-compared with parathyroid hormone (PTH) for detecting high or low bone turnover in CKD. A systematic review and meta-analysis of observational studies. Adult patients with CKD (all stages). Studies were included if they compared serum bone turnover markers to bone histomorphometry. Two reviewers independently extracted study characteristics, population details, assay types, cutoff thresholds, sensitivity, specificity, and area under the ROC curve (AUC). A meta-analysis using random-effects models pooled AUCs for detecting high and low bone turnover per marker including intact PTH. Subgroup analyses assessed assay types (mass-based vs activity-based BALP, intact vs total PINP) and CKD stages. Eight studies were included with 6 studies providing AUC data for pooling. Intact PINP and BALP (mass assay) demonstrated the highest accuracy for identifying high bone turnover, whereas intact PTH showed only moderate discriminatory performance. For low turnover, TRAP5b, BALP, and intact PINP displayed comparable accuracy, while total PINP and PTH were less reliable. Reported diagnostic thresholds varied but generally yielded sensitivities of 50%-90% and specificities of 60%-95%. One study suggested that combining bone turnover markers, such as PINP with BALP or TRAP5b, may further enhance diagnostic discrimination. Limited representation of non-hemodialysis populations, small sample sizes, and lack of standardized methods to determine thresholds. TRAP5b, BALP, and intact PINP demonstrated moderate diagnostic accuracy and may support treatment decisions in CKD when bone biopsy is not feasible. (PROSPERO: CRD420251003120).

Accurate survival prediction is a key component to patient quality of life (QoL)-centered treatment. It allows, for example, setting reasonable treatment goals or timely palliative care referral. However, clinical estimates are often too optimistic leading to lower patient QoL before death. Patient-reported outcomes (PROs) have proven to be an important survival predictor improving the overall prognostic accuracy. In this study, transformer architecture was explored to leverage PRO trajectories to improve survival accuracy and identify the most prognostic PRO symptoms for survival prediction. We analyzed 475 (cross-validation discovery set: 380; held-out testing set: 95) early-stage non-small cell lung cancer (NSCLC) patients who underwent SBRT treatment and routinely completed the Edmonton Symptom Assessment Scale (ESAS). A transformer-based model was developed to perform longitudinal modeling of overall survival (OS), incorporating PROs collected at multiple post-treatment follow-ups, as well as clinical and demographic variables. The performance of the proposed model was compared to traditional outcome modeling approaches, including univariate and multivariate (time-varying) Cox proportional hazards regression (CoxPH) and joint probability survival modelling. The best-performing transformer model was interpreted using the SHapley Additive exPlanation (SHAP) values, and the most prognostically relevant ESAS symptoms were identified through a backward elimination procedure guided by concordance index (c-index) and area under the ROC curve (AUC). The best-performing transformer model achieved a cross-validated c-index of 0.753 [95% CI: 0.742-0.764] and an AUC of 0.862 [95% CI: 0.846-0.878] on the discovery set. On the heldout test set, the model reached a c-index of 0.694 and an AUC of 0.785, evaluated at the last time point. It significantly outperformed both Cox models and the joint probability model. Model interpretation using SHAP values and backward elimination identified appetite loss, pain, overall well-being, and shortness of breath as the most prognostically relevant symptoms for survival prediction. Transformer-based survival models that integrate longitudinal PROs significantly enhance prognostic accuracy in SBRT-treated NSCLC patients. Loss of appetite and pain emerged as the most predictive symptoms, followed by overall wellbeing and shortness of breath. These findings suggest that targeted, symptom-focused PROs tracking could streamline clinical implementation and improve survival estimation in routine oncology care.

Performance of a complete AI radiographic suite across 258,373 X-rays from 26 countries: A worldwide evaluation.