Random Survival Forest Research Articles

Using Machine Learning and Feature Importance to Identify Risk Factors for Mortality in Pediatric Heart Surgery

Background: The objective of this IRB-approved retrospective monocentric study was to identify risk factors for mortality after surgery for congenital heart defects (CHDs) in pediatric patients using machine learning (ML). CHD belongs to the most common congenital malformations, and remains the leading mortality cause from birth defects. Methods: The most recent available hospital encounter for each patient with an age <18 years hospitalized for CHD-related cardiac surgery between the years 2011 and 2020 was included in this study. The cohort consisted of 1302 eligible patients (mean age [SD]: 402.92 [±562.31] days), who were categorized into four disease groups. A random survival forest (RSF) and the ‘eXtreme Gradient Boosting’ algorithm (XGB) were applied to model mortality (incidence: 5.6% [n = 73 events]). All models were then applied to predict the outcome in an independent holdout test dataset (40% of the cohort). Results: RSF and XGB achieved average C-indices of 0.85 (±0.01) and 0.79 (±0.03), respectively. Feature importance was assessed with ‘SHapley Additive exPlanations’ (SHAP) and ‘Time-dependent explanations of machine learning survival models’ (SurvSHAP(t)), both of which revealed high importance of the maximum values of serum creatinine observed within 72 h post-surgery for both ML methods. Conclusions: ML methods, along with model explainability tools, can reveal interesting insights into mortality risk after surgery for CHD. The proposed analytical workflow can serve as a blueprint for translating the analysis into a federated setting that builds upon the infrastructure of the German Medical Informatics Initiative.

Transferability and comparability of sewer deterioration models – a case study on Norwegian sewer data

ABSTRACT The adoption of emerging machine learning (ML) techniques in sewer deterioration modelling remains limited. This is primarily due to a lack of comparisons with popular models to evaluate if they perform superiorly and the extensive datasets required for ML training, which many small- to medium-sized municipalities do not possess. This study compares an emerging ML technique, random survival forest (RSF), with two established models: support vector machine (SVM) and GompitZ. Additionally, it examines the transferability of RSF models to different municipalities. The results indicate that both RSF and GompitZ have unique strengths and weaknesses, suggesting parallel use with statistical models. RSF’s performance was comparable to, or better than, SVM. Furthermore, globally trained RSF models, which use aggregated datasets of multiple municipalities, perform similarly to locally trained models, which are trained on representative municipal data. However, globally trained models are promising for municipalities lacking data or expertise to develop their own models.

Comparison of multiple machine learning models for predicting prognosis of pancreatic ductal adenocarcinoma based on contrast-enhanced CT radiomics and clinical features

ObjectiveThe aim of this study was to evaluate the prognostic potential of combining clinical features and radiomics with multiple machine learning (ML) algorithms in pancreatic ductal adenocarcinoma (PDAC).MethodsA total of 116 patients with PDAC who met the eligibility criteria were randomly assigned to a training or validation cohort. Seven ML algorithms, including Supervised Principal Components, stepwise Cox, Random Survival Forest, CoxBoost, Least absolute shrinkage and selection operation (Lasso), Ridge, and Elastic network, were integrated into 43 algorithm combinations. Forty-three radiomics models were constructed separately using radiomics features extracted from arterial phase (AP), venous phase (VP), and combined arterial and venous phase (AP+VP) images. The concordance index (C-index) of each model was calculated. The model with the highest mean C-index was identified as the best model for calculating the radiomics score (Radscore). Univariate and multivariate Cox analyses were used to identify independent prognostic indicators and create a clinical model for prognosis prediction. The multivariable Cox regression was used to combine Radscore with clinical features to create a combined model. The efficacy of the model was evaluated using the C-index, calibration curves, and decision curve analysis (DCA).ResultsThe model based on the Lasso+StepCox[both] algorithm constructed using AP+VP radiomics features showed the best predictive ability among the 114 radiomics models. The C-indices of the model in the training and validation cohorts were 0.742 and 0.722, respectively. Based on the results of the univariate and multivariate Cox regression analyses, sex, Tumor-Node-Metastasis (TNM) stage, and systemic inflammation response index were included to build the clinical model. The combined model, incorporating three clinical factors and AP+VP-Radscore, achieved the highest C-indices of 0.764 and 0.746 in the training and validation cohorts, respectively. In terms of preoperative prognosis prediction for PDAC, the calibration curve and DCA showed that the combined model had a good consistency and greatest net benefit.ConclusionA combined model of clinical features and AP+VP-Radscore screened using multiple ML algorithms has an excellent ability to predict the prognosis of PDAC and may provide a noninvasive and effective method for clinical decision-making.

Abstract 4129397: Cardiovascular Risk Models Using Large-Scale Physical Examination Data

Background: Physical examination (PE) data offer more potential key risk factors that can enhance the development of risk models. Additionally, the latest CVD risk model, PREVENT, developed from an American population, necessitates validation within China. Objectives: This study aimed to develop CVD risk models based on large-scale PE records and compare them with PREVENT model in Chinese adults. Methods: Data were collected from a China national database (2016-2024), including 929682 participants aged 30-79 years without previous CVD (Table I). The variables covered demographics, comorbidities, and PE items, resulting in 59 features as input. The endpoint events were defined as total CVD and its subtypes, ASCVD and heart failure (HF). Sex-specific survival analysis was developed based on age-scale, utilizing Cox Proportional Hazards and Random Survival Forests (RSF). Feature importance was assessed using the average minimal depth in RSF. The performance was evaluated by the Harrell C-statistic in the validation set. Results: The results showed that our RSF model achieved the best performance in total CVD, with the highest C-index of 0.82 (Table II). The top 15 predictive risk factors for each gender are presented in Figure 1. Additionally, validation of the PREVENT model in our Chinese population dataset showed its great generalizability, particularly for HF events. Conclusions: Overall, this research underscores the importance of leveraging large-scale PE data and advanced modeling techniques to improve CVD risk prediction. Beyond traditional risk factors, greater emphasis should be placed on more demographic characteristics, comorbidities, urinalysis, and ECG features.

Abstract 4143270: PREDICTIVE MODELS AID PHYSICIAN PROGNOSTICATION: A SECONDARY ANALYSIS EVALUATING INTEGRATED MODEL AND PHYSICIAN PROGNOSTIC ESTIMATES IN PATIENTS WITH HEART FAILURE WITH REDUCED EJECTION FRACTION

Background: In recent studies from a multicenter Canadian cohort of outpatients with heart failure (HF), we found that model predictions were significantly more accurate than HF cardiologists. In this study, trying to mimic practice, we evaluated the additional predictive value and clinical impact of model predictions to refine physician estimated risk of 1-year mortality by combining model and physician estimates. Methods: We included consented consecutive HF outpatients (LVEF <40%) followed at 11 HF clinics in Canada. HF cardiologists estimated patient 1-year mortality using their clinical judgment. We calculated model predicted mortality using the Seattle HF Model (SHFM). We followed patients for at least a year to record mortality (or urgent heart transplant or ventricular assist device implant as mortality-equivalent events). Using random forest survival model and cross-validation, we compared the performance SHFM and the HF cardiologist alone, and the integrated HF cardiologist and the SHFM predictions by evaluating model discrimination (c-statistic), calibration (observed vs predicted event rate), risk reclassification and clinical net benefit analyses. Results: Among 1,643 HF patients, 1-year event rate was 9% (95%CI 8%-11%). The SHFM had the adequate discrimination (c-statistic 0.76) and excellent calibration while cardiologists showed adequate discrimination (c-statistic 0.75) and poor calibration with significant risk overestimation ( Figure 1 ). When the SHFM estimates were added physician predictions, discrimination significantly improved (0.82, 95%CI 0.78-0.86) with excellent calibration. By risk reclassification analysis, among patients with events, HF cardiologist better reclassified 44% than the SHFM or the integrated model. Among patients without event, however, HF cardiologists worse risk-classified 52% in comparison to SHFM and 71% to the integrated model. By net clinical benefit analysis ( Figure 2 ), when the decision to treat involves patients with 1-year mortality of >5%, SHFM predictions would lead to higher benefit than guiding care by physician judgement. Integrating model and HF cardiologist predictions led to minimally increased benefit in comparison to SHFM alone. Conclusions: Integrating prediction from the SHFM to physician judgment or using the SHFM alone showed superior accuracy than HF cardiologist predictions, proving that model-informed care may provide more accurate prognostic information to tailor clinical decision making.

SPINK5 is a key regulator of eosinophil extracellular traps in head and neck squamous cell carcinoma

Enhanced infiltration of eosinophils is observed surrounding solid tumors. Some studies indicate that Eosinophil extracellular traps (EETs) play a crucial role in tumor progression and metastasis. However, its specific role in head and neck squamous cell carcinoma (HNSCC) remains unclear. This study established a gene set associated with eosinophil differentiation, chemotaxis, and EETs release from previous research. Employing bioinformatics techniques, the expression and biological significance of these genes in HNSCC were analyzed. Briefly, unsupervised clustering based on expression patterns of 133 EETs-related genes to classify TCGA-HNSCC patients. Immune cell infiltration patterns were assessed using “ImmuCellAI” package. A prognostic model was constructed using ten algorithms, with EETs-related gene sets as input features. Here, unsupervised clustering of samples into two types revealed worse prognosis for Cluster 1 (C1) patients after the first year. Cluster 2 (C2) exhibited higher ImmuneScore, but with a distinct immune cell infiltration pattern from the C1. Additionally, high eosinophil abundance only in the C2 had a positive prognostic impact. Serine peptidase inhibitor kazal type 5 (SPINK5) emerged as a potential key gene mediating the formation of EETs in HNSCC. EETs not only exhibit a positive correlation with diverse anti-cancer pathways but also demonstrate positive associations with processes such as proliferation, migration, and other critical pathways. The random survival forest (RSF) model was identified as the optimal eosinophil-related prognostic model. Collectively, this study elucidates the potential impact and mediating pathways of EETs on tumors, providing a reference for targeted therapy based on EETs-related genes.

Liquid-Liquid Phase Separation in the Prognosis of Lung Adenocarcinoma: An Integrated Analysis.

Lung adenocarcinoma (LUAD) is a highly lethal malignancy. Liquid-Liquid Phase Separation (LLPS) plays a crucial role in targeted therapies for lung cancer and in the progression of lung squamous cell carcinoma. However, the role of LLPS in the progression and prognosis of LUAD remains insufficiently explored. This study employed a multi-step approach to identify LLPS prognosis-related genes in LUAD. First, differential analysis, univariate Cox regression analysis, Random Survival For-est (RSF) method, and Least Absolute Shrinkage and Selection Operator (LASSO) Cox regres-sion analysis were utilized to identify five LLPS prognosis-related genes. Subsequently, LASSO Cox regression was performed to establish a prognostic score termed the LLPS-related progno-sis score (LPRS). Comprehensive analyses were then conducted based on the LPRS, including survival analysis, clinical feature analysis, functional enrichment analysis, and tumor microenvi-ronment assessment. The LPRS was integrated with additional clinicopathological factors to de-velop a prognostic nomogram for LUAD patients. Immunohistochemical validation was per-formed on clinical tissue samples to further validate the findings. Finally, the relationship be-tween KRT6A, one of the identified genes, and epidermal growth factor receptor (EGFR) muta-tions was investigated. The LPRS was established using five LLPS-related genes: IGF2BP1, KRT6A, LDHA, PKP2, and PLK1. Higher LPRS was closely associated with poor survival outcomes, gender, progression-free survival (PFS), and advanced TNM stage. Furthermore, LPRS emerged as an independent prognostic factor for LUAD. A nomogram integrating LPRS, TNM stage, and age demonstrated remarkable predictive accuracy for prognosis among patients with LUAD. LLPS likely influences LUAD prognosis through the activity of IGF2BP1, KRT6A, LDHA, PKP2, and PLK1. KRT6A exhibits significant upregulation in LUAD, particularly in patients with EGFR mutations. This study introduces a novel LPRS model that demonstrates high accuracy in pre-dicting the clinical prognosis of LUAD. Moreover, the findings suggest that KRT6A may play a critical role in the LLPS-mediated malignant progression of LUAD.

Random survival forest algorithm for risk stratification and survival prediction in gastric neuroendocrine neoplasms.

This study aimed to construct and assess a machine-learning algorithm designed to forecast survival rates and risk stratification for patients with gastric neuroendocrine neoplasms (gNENs) after diagnosis. Data on patients with gNENs were extracted and randomly divided into training and validation sets using the Surveillance, Epidemiology, and End Results database. We developed a prediction model using 10 machine learning algorithms across 101 combinations to forecast cancer-related mortality in patients with gNENs, selecting the best model using the highest mean over a sequence of time-dependent area under the receiver operating characteristic (ROC) curve (AUC). The performance of the final model was assessed through time-dependent ROC curves for discrimination and calibration curves for calibration. The maximum selection rank method was used to determine the best prognostic risk score threshold for classifying patients into high- and low-risk groups. Afterward, Kaplan-Meier analysis and log-rank test were used to compare survival rates among these groups. Our study examined 775 patients with gNENs, dividing them into training and validation sets. A training set comprised 543 patients, with a median follow-up of 42 months and cumulative mortality rates of 40.0% at 1 year, 48.6% at 3 years, and 54.0% at 5 years. A validation set comprised 232 patients, with cumulative mortality rates of 29.1% at 1 year, 43.5% at 3 years, and 53.2% at 5 years. The optimal random survival forest (RSF) model (mtry = 4, node size = 5) achieved an AUC of 0.839 for survival prediction in the training set. Comprising 11 variables such as demographics, treatment details, tumor characteristics, T staging, N staging, and M staging, the RSF model revealed high predictive accuracy with AUCs of 0.92, 0.96, and 0.96 for 1-, 3-, and 5-year survival, respectively, which was consistently reflected in the validation set with AUCs of 0.88, 0.92, and 0.89, respectively. Moreover, patients were risk-stratified. Although our RSF model effectively stratified patients into different prognostic groups, it needs external validation to confirm its utility for noninvasive prognostic prediction and risk stratification in gNENs. Further research is required to verify its broader clinical applicability.

Prediction of recurrence-free survival and risk factors of sinonasal inverted papilloma after surgery by machine learning models

ObjectivesOur research aims to construct machine learning prediction models to identify patients proned to recurrence after inverted papilloma (IP) surgery and guide their follow-up treatment.MethodsThis study collected 210 patients underwent IP resection surgery at a university hospital from January 2010 to December 2023. Six machine learning algorithms including ExtraSurvivalTrees (EST), GradientBoostingSurvivalAnalysis (GBSA), RandomSurvivalForest (RSF), SurvivalSVM, Coxnet and Coxph, were used to construct the prediction models. Shapley Additive Explanations (SHAP) values were used to explain the importance of various features in predicting IP recurrence.ResultsWe found that the recurrence rate of IP patients is 20.00%, with a median recurrence time of 35.5 months. Multivariate Cox regression analysis identified mild or moderate dysplasia as an independent risk factor for recurrence. The EST model performs the best in predicting postoperative recurrence of IP, with C-index of 0.968 and 0.878 in the training and testing sets. SHAP emphasizes five important predictive factors for recurrence, including bone defects, orbital involvement, smoking, no processing of tumor attachment sites and drinking.ConclusionsTo our knowledge, this is the first study to use multiple ML models to predict postoperative recurrence of IP. The EST model has the best predictive performance, with SHAP emphasizing several key predictive factors for IP recurrence. This study emphasizes the practicality of machine learning algorithms in predicting IP clinical outcomes, providing valuable insights into the potential for improving clinical decision-making.

Balancing accuracy and Interpretability: An R package assessing complex relationships beyond the Cox model and applications to clinical prediction

BackgroundAccurate and interpretable models are essential for clinical decision-making, where predictions can directly impact patient care. Machine learning (ML) survival methods can handle complex multidimensional data and achieve high accuracy but require post-hoc explanations. Traditional models such as the Cox Proportional Hazards Model (Cox-PH) are less flexible, but fast, stable, and intrinsically transparent. Moreover, ML does not always outperform Cox-PH in clinical settings, warranting a diligent model validation. We aimed to develop a set of R functions to help explore the limits of Cox-PH compared to the tree-based and deep learning survival models for clinical prediction modelling, employing ensemble learning and nested cross-validation. MethodsWe developed a set of R functions, publicly available as the package “survcompare”. It supports Cox-PH and Cox-Lasso, and Survival Random Forest (SRF) and DeepHit are the ML alternatives, along with the ensemble methods integrating Cox-PH with SRF or DeepHit designed to isolate the marginal value of ML. The package performs a repeated nested cross-validation and tests for statistical significance of the ML’s superiority using the survival-specific performance metrics, the concordance index, time-dependent AUC-ROC and calibration slope.To get practical insights, we applied this methodology to clinical and simulated datasets with varying complexities and sizes. ResultsIn simulated data with non-linearities or interactions, ML models outperformed Cox-PH at sample sizes ≥ 500. ML superiority was also observed in imaging and high-dimensional clinical data. However, for tabular clinical data, the performance gains of ML were minimal; in some cases, regularised Cox-Lasso recovered much of the ML’s performance advantage with significantly faster computations. Ensemble methods combining Cox-PH and ML predictions were instrumental in quantifying Cox-PH’s limits and improving ML calibration. Traditional models like Cox-PH or Cox-Lasso should not be overlooked while developing clinical predictive models from tabular data or data of limited size. ConclusionOur package offers researchers a framework and practical tool for evaluating the accuracy-interpretability trade-off, helping make informed decisions about model selection.

Delta-radiomics Approach Using Contrast-enhanced and Non-contrast-enhanced Computed Tomography Images for Predicting Distant Metastasis in Patients with Borderline Resectable Pancreatic Carcinoma

PurposeTo predict distant metastasis (DM) in patients with borderline resectable pancreatic carcinoma (BRPC) using delta-radiomics features calculated from contrast-enhanced computed tomography (CECT) and non-CECT images. MethodsAmong 250 patients who underwent radiotherapy at our institution between February 2013 and December 2021, 67 patients were deemed eligible. A total of 11 clinical features and 3,906 radiomics features were incorporated. Radiomics features were extracted from CECT and non-CECT images and the differences between these features were calculated, resulting in delta-radiomics features. The patients were randomly divided into the training (70%) and test (30%) datasets for model development and validation. Predictive models were developed with clinical features (clinical model), radiomics features (radiomics model), and a combination of the abovementioned features (hybrid model) using Fine–Gray regression (FG) and random survival forest (RSF). Optimal hyperparameters were determined using stratified five-fold cross-validation. Subsequently, the developed models were applied to the remaining test datasets, and the patients were divided into the high- or low-risk groups based on their risk scores. Prognostic power was assessed using the concordance index (C-index) with 95% confidence intervals obtained through 2,000 bootstrapping iterations. Statistical significance between the two groups was assessed using Gray's test. ResultsAt a median follow-up period of 23.8 months, 47 (70.1%) patients developed DM. The C-indices of the FG-based clinical, radiomics, and hybrid models were 0.548, 0.603, and 0.623, respectively, in the test dataset, whereas those of the RSF-based models were 0.598, 0.680, and 0.727, respectively. The RSF-based model, including delta-radiomics features, significantly divided the cumulative incidence curves into two groups (P<0.05). The feature map of the gray-level size-zone matrix showed that the difference in feature values between CECT and non-CECT images correlated with the incidence of DM. ConclusionsDelta-radiomics features obtained from CECT and non-CECT images using RSF successfully predict the incidence of DM in patients with BRPC.

Integrated multi-level omics profiling of disulfidptosis identifis SPAG4 as an innovative immunotherapeutic target in glioblastoma.

To investigate the association between disulfidptosis-related genes (DFRGs) and patient prognosis, while concurrently identifying potential therapeutic targets in glioblastoma (GBM). We retrieved RNA sequencing data and clinical characteristics of GBM patients from the TCGA database. We found there was a total of 6 distinct clusters in GBM, which was identified by the t-SNE and UMAP dimension reduction analysis. Prognostically significant genes in GBM were identified using the limma package, coupled with univariate Cox regression analysis. Machine learning algorithms were then applied to identify central genes. The CIBERSORT algorithm was utilized to assess the immunological landscape across different GBM subtypes. In vitro and in vivo experiments were conducted to investigate the role of SPAG4 in regulating the proliferation, invasion of GBM, and its effects within the immune microenvironment. 23 genes, termed DFRGs, were successfully identified, demonstrating substantial potential for establishing a prognostic model for GBM. Single cell analysis revealed a significant correlation between DFRGs and the progression of GBM. Utilizing individual risk scores derived from this model enabled the stratification of patients into two distinct risk groups, revealing significant variations in immune infiltration patterns and responses to immunotherapy. Utilizing the random survival forests algorithm, SPAG4 was identified as the gene with the highest prognostic significance within our model. In vitro studies have elucidated SPAG4's significant role in GBM pathogenesis, potentially through the regulation of fatty acid metabolism pathways. Our in vivo investigations using a subcutaneous xenograft model have confirmed SPAG4's influence on tumor growth and its capacity to modulate the immune microenvironment. Advanced research hints that SPAG4 might achieve immune evasion by increasing CD47 expression, consequently reducing phagocytosis. These findings highlight SPAG4 as a potential GBM therapeutic target and emphasize the complexity of the immune microenvironment in GBM progression.

Predicting survival benefits of immune checkpoint inhibitor therapy in lung cancer patients: a machine learning approach using real-world data.

Due to the heterogeneity in the effectiveness of immunotherapy for lung cancer, identifying predictors is crucial. This study aimed to develop a machine learning model to identify predictors of overall survival in lung cancer patients treated with immune checkpoint inhibitors (ICIs). A retrospective analysis was performed on data from 1314 lung cancer patients at the Chongqing University Cancer Hospital from September 2018 to September 2022. We used the random survival forest (RSF) model to identify survival-influencing factors, using backward elimination for variable selection. A Cox proportional hazards (CPH) model was constructed using the most significant predictors. We assessed model performance and generalizability using time-dependent receiver operating characteristics (ROC) and predictive error curves. The RSF model demonstrated better predictive accuracy than the CPH (IBS 0.17 vs. 0.17; C-index 0.91 vs. 0.68), with better discrimination and prediction performance. The influential variables identified included D-dimer, Karnofsky performance status, albumin, surgery, TNM stage, platelet count, and age. The RSF model, which incorporated these variables, achieved area under the curve (AUC) scores of 0.95, 0.94, and 0.98 for 1-, 3-, and 5-year survival predictions, respectively, in the training set. The validation set showed AUCs of 0.94, 0.90, and 0.95, respectively, exceeding the performance of the CPH model. The study successfully developed a machine learning model that accurately predicted the survival benefits of ICI therapy in lung cancer patients, supporting clinical decision-making in lung cancer treatment.

AI derived myocardial left ventricular volumes outperforms clinician segmentation for prediction of all-cause mortality

Abstract Background Measuring heart structure and function drives much cardiac decision making and therapy. Precise quantification is therefore of high importance, but manual (clinician) segmentation introduces measurement variability. Purpose Automated segmentation of cardiac MRI (CMR) left ventricular (LV) volumes provides improved precision but how this translates into clinical outcomes has yet to be defined. In order to test predictive accuracy in a wide range of adverse LV morphologies, we applied AI LV segmentation to patients with coronary disease or cardiomyopathy. Methods Scans were analysed from a retrospective clinical CMR database and mortality data was obtained from a national database. Clinician measurement of indexed LV end-diastolic volume (LVEDVi), LV mass (LVMi), myocardial contraction fraction (MCF: ratio of stroke volume to myocardial volume; a measure of myocardial efficiency) and EF were recorded from clinical reports (all senior clinicians with &amp;gt;5 years level 3 experience) and compared to AI measurement using a segmentation algorithm. This has previously been shown to exceed human precision and required no manual correction (1,2). Random survival forests were built to identify overall predictive value of AI vs clinician derived LV volumes. These are machine-learning algorithms for predicting time-to-event outcomes and can capture complex relationships between predictors and outcome. Random survival forests were built from AI and clinician models with EF, EDVi, LVMi and MCF as predictors for all-cause mortality. A train:test (80:20) split, stratified for outcome was employed. Concordance (C)-indices (measuring predictive accuracy) and permutation importances (measuring significance of each predictor within the model) (Table) with 1000 bootstrapped 95% confidence intervals and p-values were derived. Results 8,299 patients were analysed. Patients were 60(51-70) years old and 37% female. 49% had coronary disease, 25% had DCM and 19% had HCM. At a median follow-up of 5.3 years (IQR 4.2-6.6 years) there were 1,384 deaths (17%). AI derived LV-parameter models had higher predictive accuracy than clinician derived models (C-Indices: 0.66±0.002 vs 0.65±0.005, p&amp;lt;0.001). A subgroup analysis of patients undergoing ischaemia testing for coronary disease was performed (n=3,560) and markers of ischaemia (positive vs negative) and infarction (non-viable vs viable) were included in the models. AI had a higher predictive accuracy compared to clinicians when including these additional predictors (C-Indices: 0.68±0.03 vs 0.63±0.01, p&amp;lt;0.001). Conclusion AI derived LV volumes outperforms clinicians in prediction of all-cause mortality reflecting improved precision. Improvements in predictive accuracy are likely to be clinically important when applied to populations and high-risk groups. Manual clinician segmentation should be superseded by clinician-supervised AI.

Comprehensive machine learning models for prediction of heart failure in 476,393 women and men from the UK Biobank reveal sex differences and underutilized risk factors

Abstract Background Machine learning (ML) models provide potential advantage over ‘traditional’ regression models in heart failure (HF) prediction. Objective To compare performances of Cox PH models and ML survival models for incident HF in men and women without prevalent ischemic heart disease (IHD). We also aimed to identify potential high-risk precursors otherwise ignored by conventional survival models, and to investigate differences between sex-specific models. Methods We included 476,393 participants (55.6% women) from the UK Biobank, after excluding participants with a history of HF or IHD, and defined sex-specific datasets. We predicted incident HF events using over 400 baseline characteristics. We constructed multivariable Cox PH models, which included all predictor variables and subsequently only those remaining after LASSO stability selection. We also developed two supervised ML models (Random Survival Forest (RSF), eXtreme Gradient Survival Boosting (XGBoost)). We identified the 15 most important sex-specific predictors in each model and performances were compared using the C-index. Models were validated using hold-out sets. Results During 12.3 ± 1.9 years of follow-up, 4680 (1.76%) women and 6631 (3.14%) men developed incident HF. XGBoost showed the best performance during model training (C-index, training: 0.89 in men, 0.97 in women; validation 0.77 in men, 0.80 in women). The multivariable Cox model performed second-best (C-index, training: 0.78 in men, 0.82 in women; validation: 0.76 in men, 0.78 in women). RSF performed slightly worse (C-index, training: 0.75 in men, 0.79 in women; validation: 0.75 in men, 0.79 in women) but did not show performance drop during validation. LASSO stability selection performed similar to RSF. Age, self-reported lifetime treatments and medications, cystatin-C, waist circumference and FEV1-scores were identified as strong risk factors in all models for both sexes. Reduced albumin levels and elevated HbA1c were more strongly associated with high risk in men, while elevated systolic BP showed higher importance in women. Traditional Cox models observed CRP as important only in men, while the ML models identified CRP as important for both sexes. Neutrophil count was considered a strong risk factor in both sexes in the traditional Cox models, yet it was not among the most important predictors in both ML models. Presence of other heart disease (which included a.o. pericardial disease, valve disorders and arrhythmias) was an important predictor variable only in the ML models. Conclusion ML models showed similar performance to Cox PH models for HF prediction. Despite this, differences in predictor importance were identified between models. Sex-specific risk predictors were found, and FEV1 score, which is not commonly included in existing models, was identified as an important risk factor. These results suggest that ML models may reveal additional insights that would otherwise remain unnoticed.

Improving prognostic accuracy in ischemic cardiomyopathy: the CMR-LGE score

Abstract Background There is a need for accurate prognostic stratification tools in patients with ischemic cardiomyopathy (ICM). Several studies have shown the considerable impact of cardiac magnetic resonance (CMR) imaging findings notably late gadolinium enhancement (LGE) but an easy interpretable score to guide clinical practice is lacking. Purpose To determine the CMR parameters most predictive of mortality in a cohort of ICM patients with left ventricular ejection fraction (LVEF) less than 50%, and then to develop a readily interpretable score based on these parameters. Methods Between 2008 and 2022, all consecutive patients with a history of ICM with LVEF&amp;lt;50% and presence of ischemic LGE on viability CMR examination, were recruited by two independent centers. The primary outcome was all-cause death using French National Registry of Death. The first center (N=2,900) was designated as the derivation cohort for variable selection and score development, and the second center (N=691) was used as the validation cohort to evaluate the performance of the final score. Thirty-two variables (17 clinical and 15 CMR) were initially assessed. Automatic feature selection was performed using a machine-learning approach with a Random Survival Forest (RSF) algorithm. Using the selected variables, our proposed score, the CMR-LGE score, was derived from coefficients of the Cox regression. Performance evaluation on the validation cohort was conducted using Harrel’s C index compared with traditional prognostic factors associated with poor outcomes in ICM. Categories for the prognostic score were identified using a survival conditional inference tree analysis on the derivation cohort, aiming to maximize the log-rank score. Results Among the 3,591 patients included (mean age 65±12 years; 75% male; mean LVEF 44±6%), 549 (15.3%) died over a median follow-up of 9 (interquartile range: 7–12) years. Feature selection with RSF highlighted that the most important variables to predict death were LGE variables: the extent, the location (septal and anterior) and the transmurality of ischemic LGE as well as the extent of additional midwall LGE (Figure 1). Following these findings, we developed the CMR-LGE score by rounding the coefficient of a Cox regression (Figure 2A). Harrel’s C index of CMR-LGE score outperformed traditional prognostic factors, including LVEF (0.88 vs 0.69). Finally, based on our CMR-LGE score, we identified a low-risk population (CMR-LGE score below 6) and a high-risk population (CMR-LGE above 8), validated using survival curves in the validation cohort (Figure 2B). Conclusion Using RSF, we identified that the 5 most important CMR variables to predict mortality in ICM were all LGE features. Our CMR-LGE score showed excellent performance to stratify patient risk compared to traditional prognostic factors including LVEF.Variable selectionModel building and evaluation

Multi-modal artificial intelligence-based prediction for cardiovascular mortality after transcatheter aortic valve implantation: a time-to-event survival prediction

Abstract Background Predicting long-term outcomes following Transcatheter Aortic Valve Implantation (TAVI) presents a significant challenge. Objective We aimed to predict post-TAVI long-term cardiovascular mortality using multimodal data and employing time-to-event artificial intelligence (AI) algorithms. Methods This study enrolled 3,973 consecutive patients from the prospective TAVI registry (between August 2007 and June 2023). Multimodal information was collected and extracted as numerical data for each patient, encompassing clinical history, medication, laboratory, procedural characteristics, ECG, angiography, echocardiography (TTE and TEE), and CT. Clinical outcomes were adjudicated by an independent clinical event committee. From the multimodal information, 147 features were extracted at baseline before the TAVI procedure. The datasets were divided into a training/validation set (80%) and a hold-out test set (20%) for model development and evaluation, respectively. Cox-LASSO feature selection was employed to identify the most relevant features, followed by the implementation of Random Survival Forest (RSF), Gradient Boosting Survival Analysis (GBSA), and Survival Support Vector Machine (SSVM) to predict the timing and occurrence of cardiac death events in patients. Feature selection, model parameters, and hyperparameters selection were performed using 10-fold cross-validation on the training set. Model performance was evaluated using different metrics, such as the concordance index (c-index) and cumulative Area Under the Curve (C-AUC) on a 20% untouched test set. Results In the patient cohort, 28.4% experienced cardiovascular death. Features were selected from various modalities, including 22 from clinical history, two from medication, eight from laboratory tests, one from ECG, eight from CT scans, and five from echocardiography and angiography. In the univariate Cox-PH model, all features exhibited a C-index ranging from 0.50 to 0.57. In the test set, the conventional Multivariate Cox-PH model achieves a c-index of 0.59 (C-AUC: 0.56). RSF, GBSA, and SSVM achieve a C-index of 0.70 (C-AUC: 0.69), 0.66 (C-AUC: 0.64), and 0.69 (C-AUC: 0.68), respectively, in the holdout test set. Conclusion Integrating multimodal information, encompassing imaging, signal, laboratory data, and clinical history, with AI algorithms has markedly enhanced the prediction of long-term cardiac outcomes following the TAVI procedure. The initially developed model can predict the probability of cardiac death up to 10 years. Furthermore, the RFS model significantly outperformed the conventional Cox model in predicting time-to-event outcomes.

Prediction model for survival of younger patients with breast cancer using the breast cancer public staging database

Breast cancer (BC) is a major contributor to female mortality worldwide, particularly in young women with aggressive tumors. Despite the need for accurate prognosis in this demographic, existing studies primarily focus on broader age groups, often using the SEER database, which has limitations in variable selection. This study aimed to develop an ML-based model to predict survival outcomes in young BC patients using the BC public staging database. A total of 3,401 patients with BC were included in the study. Patients were categorized as younger (n = 1574) and older (n = 1827). We applied several survival models—Random Survival Forest, Gradient Boosting Survival, Extra Survival Trees (EST), and penalized Cox models (Lasso and ElasticNet)—to compare mortality characteristics. The EST model outperformed others in predicting mortality for both age groups. Older patients exhibited a higher prevalence of comorbidities compared to younger patients. Tumor stage was the primary variable used to train the model for mortality prediction in both groups. COPD was a significant variable only in younger patients with BC. Other variables exhibited varying degrees of consistency in each group. These findings can help identify high-risk young female patients with BC who require aggressive treatment by predicting the risk of mortality.

Detection and risk stratification of cardiac amyloidosis patients by integration of imaging and non-imaging data using a machine learning approach

Abstract Background With the advent of amyloid-targeting therapies, early and reliable diagnosis as well as precise risk estimation of cardiac amyloidosis (CA) have become of substantial importance. While the current diagnostic approach relies on difficult-to-standardize, visual interpretation of 99mTc-scintigraphy, risk assessment is largely based on a combination of blood and imaging parameters. Purpose Aim of this study is to assess the possibility of machine learning-based integration of scintigraphy, echocardiography, and non-imaging parameters such as blood parameters and comorbidities for the detection and risk stratification of patients with CA. Methods This study included all patients who underwent 99mTc-scintigraphy at a university affiliated tertiary care centre between 2010 and 2023. Two machine learning models were developed, and the relative predictive importance of their parameters were assessed (Figure 1): First, a model to detect patients with CA-suggestive uptake (Perugini grade 2/3) on 99mTc-scintigraphy scans; Second, a model to assess the risk of patients with CA-suggestive uptake for future heart failure hospitalization (HFH). A total of 58 features were extracted from electronic health records including blood, echocardiographic, demographic parameters and comorbidities. Scintigraphy imaging features were extracted from the raw imaging data using a deep learning approach. For the time to HFH prediction, a random survival forest machine learning model was employed. Results Overall, 12 380 consecutively enrolled patients were included, 279 (2.3%) of which were affected by CA-suggestive uptake. The machine learning model showed higher accuracy for the detection of CA (AUC 0.96 [95% CI 0.95-0.97], sensitivity 0.78 [95% CI 0.77-0.80] and specificity 0.99 [95% CI 0.99-0.99]. In the outcome analysis, the machine model showed good accuracy in predicting future HFH (C-index 0.71 [95% CI 0.68-0.75]). Parameter importance for the time to event analysis revealed right ventricular diameter, previous diagnosis of chronic heart failure and creatine kinase as the three most important predictors (Figure 2). Conclusions Detection of patients with cardiac amyloidosis and their risk estimation for heart failure can be aided using machine learning-based integration of imaging and non-imaging data. Our findings may guide novel approaches for assessing disease progression and treatment response.

Left atrial function for predicting atrial fibrillation: a machine learning approach

Abstract Background Common risk scores for atrial fibrillation (AF) are based on clinical risk factors, and little is known about the contribution of left atrial (LA) function parameters to risk assessment. Purpose To determine whether a machine learning-based model that includes LA function parameters outperforms a commonly used risk scores, the CHARGE-AF, in predicting AF. Methods Participants in the Atherosclerosis Risk in Communities (ARIC) Study (a community-based cohort study in the USA) who attended visit 5 (2011-2013) and without a history of AF were included in this analysis. Incident AF was ascertained from ECG during study visits, hospitalization discharge codes, and death certificates. Random survival forest methodology was used to build a prediction model for AF; candidate variables included traditional clinical risk factors and measures of LA structure and function. Dataset was randomly split into a training (75%) and testing (25%) cohort. For each variable, importance was measured based on the reduction in Harrell’s c-statistic if the variable was randomly permuted prior to obtaining predicted values. Variables significantly associated with AF were included in the final model. Model calibration was tested using the Greenwood-Nam-D'Agostino (GND) test. Discrimination ability of our model was assessed in the testing cohort using Harrell’s c-statistic and compared with that of the CHARGE-AF model. Results 4,909 participants (59% female, 22% Black, mean age 75 years) were included. During a median follow-up of 7.0 years (1st-3rd quartile: 5.4-7.8), 650 participants developed AF. Variables significantly associated with incident AF were race, triglycerides, systolic blood pressure, coronary heart disease, age, LA maximal volume index, NT-proBNP, and LA reservoir strain (Figure 1). LA reservoir strain was the variable most strongly associated with AF. The final model showed good discrimination (c-statistic (95% CI): 0.734 (0.698-0.768)). The GND test did not detect statistically significant miscalibration (GND p=0.08). Compared with our model, CHARGE-AF had lower discrimination ability (c-statistic (95% CI): 0.635 (0.594-0.677)) and the difference was statistically significant (difference in c-statistic (95% CI): 0.099 (0.043-0.152)). Conclusion A machine learning-based model with few variables including LA function outperforms a commonly used risk model in predicting AF. Incorporation of LA functional assessment may help inform clinicians regarding individual risk of AF.