Predictive Ability Of Machine Learning Models For Cardiogenic Shock Mortality: A Systematic Review And Meta-analysis
HFSA ePoster Library. Mylavarapu M. 10/11/26; 4234862; 256
Dr. Maneeth Mylavarapu
REGULAR CONTENT
Login now to access Regular content available to all registered users.
Abstract
Discussion Forum (0)
Rate & Comment (0)
General ePoster - EP.6.B: On 2026-10-11 13:15 (Monitor 46)
Topic: Cardiogenic Shock, Temporary Mechanical Circulatory Support, and Critical Care Cardiology
Maneeth Mylavarapu1,2, Madiha Kiyani3, Niharika Tanwar4, Neel A. Doshi5, Vaibhav Vats6, Nithin Karnan7, Farjahan Chowdhury8, Vikash Jaiswal1, Mina Kerolos1,2. 1Endeavor Health Cardiovascular Institute, Glenview, IL; 2University of Chicago Pritzker School of Medicine, Glenview, IL; 3MedStar Health, Baltimore, MD; 4Advocate Illinois Masonic Medical Center, Chicago, IL; 5Mobile Infirmary Medical Center, Mobile, AL; 6Jacobi Medical Center, Albert Einstein College of Medicine, Bronx, NY; 7MercyOne North Iowa Medical Centre, Mason City, IA; 8San Antonio Regional Hospital, Upland, CA
Introduction:
In cardiogenic shock (CS) management, early and accurate risk stratification is important in guiding clinical decisions, including the timely escalation of mechanical circulatory support (MCS) and the allocation of intensive care resources. While traditional prognostic scores are widely utilized, they often fail to capture the complex, dynamic physiological variables inherent to CS. Machine learning (ML) offers a promising alternative by synthesizing high-dimensional clinical data; however, its performance and reliability needs rigorous evaluation to justify bedside implementation.
Hypothesis:
We hypothesize that ML based predictive models will demonstrate superior prognostic accuracy and discrimination for mortality in cardiogenic shock compared to traditional clinical scoring systems.
Methods:
As per PRISMA guidelines, conducted a systematic review and meta-analysis of studies reporting the predictive accuracy of ML and traditional (TR) models for mortality in patients with CS. The primary outcome was the pooled area under the receiver operating characteristic curve (AUROC), calculated using a DerSimonian-Laird random-effects model. ML algorithms were categorized into Ensemble models, Hybrid models (including optimized and derived scores), and Others.
Results:
A total of 8 studies with 13,491 patients were included in the analysis. The pooled AUROC for ML models was 0.88 [95% CI: 0.84-0.91], while TR models achieved a pooled AUROC of 0.80 [95% CI: 0.75-0.86]. Subgroup analysis of ML internal validation revealed high, comparable predictive performance across all three subgroups i.e., Hybrid models [pooled AUROC 0.88, 95% CI: 0.84-0.93], Ensemble models [pooled AUROC 0.87, 95% CI: 0.82-0.91], and other models [pooled AUROC 0.89, 95% CI: 0.78-0.99], with overlapping confidence intervals. Although, predictive accuracy for ML models declined [pooled AUROC 0.803, 95% CI: 0.756-0.849] during external validation, this still exceeded the TR models external validation estimate [pooled AUROC 0.770, 0.702-0.801].
Conclusion: Machine learning shows promise for dynamic risk stratification in cardiogenic shock, yet its current predictive advantage over traditional models is modest. Due to significant data heterogeneity and performance decay in external validation, the clinical utility of these tools depends on future prospective, multi-center studies that prioritize standardized data quality and model interpretability.
Topic: Cardiogenic Shock, Temporary Mechanical Circulatory Support, and Critical Care Cardiology
Maneeth Mylavarapu1,2, Madiha Kiyani3, Niharika Tanwar4, Neel A. Doshi5, Vaibhav Vats6, Nithin Karnan7, Farjahan Chowdhury8, Vikash Jaiswal1, Mina Kerolos1,2. 1Endeavor Health Cardiovascular Institute, Glenview, IL; 2University of Chicago Pritzker School of Medicine, Glenview, IL; 3MedStar Health, Baltimore, MD; 4Advocate Illinois Masonic Medical Center, Chicago, IL; 5Mobile Infirmary Medical Center, Mobile, AL; 6Jacobi Medical Center, Albert Einstein College of Medicine, Bronx, NY; 7MercyOne North Iowa Medical Centre, Mason City, IA; 8San Antonio Regional Hospital, Upland, CA
Introduction:
In cardiogenic shock (CS) management, early and accurate risk stratification is important in guiding clinical decisions, including the timely escalation of mechanical circulatory support (MCS) and the allocation of intensive care resources. While traditional prognostic scores are widely utilized, they often fail to capture the complex, dynamic physiological variables inherent to CS. Machine learning (ML) offers a promising alternative by synthesizing high-dimensional clinical data; however, its performance and reliability needs rigorous evaluation to justify bedside implementation.
Hypothesis:
We hypothesize that ML based predictive models will demonstrate superior prognostic accuracy and discrimination for mortality in cardiogenic shock compared to traditional clinical scoring systems.
Methods:
As per PRISMA guidelines, conducted a systematic review and meta-analysis of studies reporting the predictive accuracy of ML and traditional (TR) models for mortality in patients with CS. The primary outcome was the pooled area under the receiver operating characteristic curve (AUROC), calculated using a DerSimonian-Laird random-effects model. ML algorithms were categorized into Ensemble models, Hybrid models (including optimized and derived scores), and Others.
Results:
A total of 8 studies with 13,491 patients were included in the analysis. The pooled AUROC for ML models was 0.88 [95% CI: 0.84-0.91], while TR models achieved a pooled AUROC of 0.80 [95% CI: 0.75-0.86]. Subgroup analysis of ML internal validation revealed high, comparable predictive performance across all three subgroups i.e., Hybrid models [pooled AUROC 0.88, 95% CI: 0.84-0.93], Ensemble models [pooled AUROC 0.87, 95% CI: 0.82-0.91], and other models [pooled AUROC 0.89, 95% CI: 0.78-0.99], with overlapping confidence intervals. Although, predictive accuracy for ML models declined [pooled AUROC 0.803, 95% CI: 0.756-0.849] during external validation, this still exceeded the TR models external validation estimate [pooled AUROC 0.770, 0.702-0.801].
Conclusion: Machine learning shows promise for dynamic risk stratification in cardiogenic shock, yet its current predictive advantage over traditional models is modest. Due to significant data heterogeneity and performance decay in external validation, the clinical utility of these tools depends on future prospective, multi-center studies that prioritize standardized data quality and model interpretability.
General ePoster - EP.6.B: On 2026-10-11 13:15 (Monitor 46)
Topic: Cardiogenic Shock, Temporary Mechanical Circulatory Support, and Critical Care Cardiology
Maneeth Mylavarapu1,2, Madiha Kiyani3, Niharika Tanwar4, Neel A. Doshi5, Vaibhav Vats6, Nithin Karnan7, Farjahan Chowdhury8, Vikash Jaiswal1, Mina Kerolos1,2. 1Endeavor Health Cardiovascular Institute, Glenview, IL; 2University of Chicago Pritzker School of Medicine, Glenview, IL; 3MedStar Health, Baltimore, MD; 4Advocate Illinois Masonic Medical Center, Chicago, IL; 5Mobile Infirmary Medical Center, Mobile, AL; 6Jacobi Medical Center, Albert Einstein College of Medicine, Bronx, NY; 7MercyOne North Iowa Medical Centre, Mason City, IA; 8San Antonio Regional Hospital, Upland, CA
Introduction:
In cardiogenic shock (CS) management, early and accurate risk stratification is important in guiding clinical decisions, including the timely escalation of mechanical circulatory support (MCS) and the allocation of intensive care resources. While traditional prognostic scores are widely utilized, they often fail to capture the complex, dynamic physiological variables inherent to CS. Machine learning (ML) offers a promising alternative by synthesizing high-dimensional clinical data; however, its performance and reliability needs rigorous evaluation to justify bedside implementation.
Hypothesis:
We hypothesize that ML based predictive models will demonstrate superior prognostic accuracy and discrimination for mortality in cardiogenic shock compared to traditional clinical scoring systems.
Methods:
As per PRISMA guidelines, conducted a systematic review and meta-analysis of studies reporting the predictive accuracy of ML and traditional (TR) models for mortality in patients with CS. The primary outcome was the pooled area under the receiver operating characteristic curve (AUROC), calculated using a DerSimonian-Laird random-effects model. ML algorithms were categorized into Ensemble models, Hybrid models (including optimized and derived scores), and Others.
Results:
A total of 8 studies with 13,491 patients were included in the analysis. The pooled AUROC for ML models was 0.88 [95% CI: 0.84-0.91], while TR models achieved a pooled AUROC of 0.80 [95% CI: 0.75-0.86]. Subgroup analysis of ML internal validation revealed high, comparable predictive performance across all three subgroups i.e., Hybrid models [pooled AUROC 0.88, 95% CI: 0.84-0.93], Ensemble models [pooled AUROC 0.87, 95% CI: 0.82-0.91], and other models [pooled AUROC 0.89, 95% CI: 0.78-0.99], with overlapping confidence intervals. Although, predictive accuracy for ML models declined [pooled AUROC 0.803, 95% CI: 0.756-0.849] during external validation, this still exceeded the TR models external validation estimate [pooled AUROC 0.770, 0.702-0.801].
Conclusion: Machine learning shows promise for dynamic risk stratification in cardiogenic shock, yet its current predictive advantage over traditional models is modest. Due to significant data heterogeneity and performance decay in external validation, the clinical utility of these tools depends on future prospective, multi-center studies that prioritize standardized data quality and model interpretability.
Topic: Cardiogenic Shock, Temporary Mechanical Circulatory Support, and Critical Care Cardiology
Maneeth Mylavarapu1,2, Madiha Kiyani3, Niharika Tanwar4, Neel A. Doshi5, Vaibhav Vats6, Nithin Karnan7, Farjahan Chowdhury8, Vikash Jaiswal1, Mina Kerolos1,2. 1Endeavor Health Cardiovascular Institute, Glenview, IL; 2University of Chicago Pritzker School of Medicine, Glenview, IL; 3MedStar Health, Baltimore, MD; 4Advocate Illinois Masonic Medical Center, Chicago, IL; 5Mobile Infirmary Medical Center, Mobile, AL; 6Jacobi Medical Center, Albert Einstein College of Medicine, Bronx, NY; 7MercyOne North Iowa Medical Centre, Mason City, IA; 8San Antonio Regional Hospital, Upland, CA
Introduction:
In cardiogenic shock (CS) management, early and accurate risk stratification is important in guiding clinical decisions, including the timely escalation of mechanical circulatory support (MCS) and the allocation of intensive care resources. While traditional prognostic scores are widely utilized, they often fail to capture the complex, dynamic physiological variables inherent to CS. Machine learning (ML) offers a promising alternative by synthesizing high-dimensional clinical data; however, its performance and reliability needs rigorous evaluation to justify bedside implementation.
Hypothesis:
We hypothesize that ML based predictive models will demonstrate superior prognostic accuracy and discrimination for mortality in cardiogenic shock compared to traditional clinical scoring systems.
Methods:
As per PRISMA guidelines, conducted a systematic review and meta-analysis of studies reporting the predictive accuracy of ML and traditional (TR) models for mortality in patients with CS. The primary outcome was the pooled area under the receiver operating characteristic curve (AUROC), calculated using a DerSimonian-Laird random-effects model. ML algorithms were categorized into Ensemble models, Hybrid models (including optimized and derived scores), and Others.
Results:
A total of 8 studies with 13,491 patients were included in the analysis. The pooled AUROC for ML models was 0.88 [95% CI: 0.84-0.91], while TR models achieved a pooled AUROC of 0.80 [95% CI: 0.75-0.86]. Subgroup analysis of ML internal validation revealed high, comparable predictive performance across all three subgroups i.e., Hybrid models [pooled AUROC 0.88, 95% CI: 0.84-0.93], Ensemble models [pooled AUROC 0.87, 95% CI: 0.82-0.91], and other models [pooled AUROC 0.89, 95% CI: 0.78-0.99], with overlapping confidence intervals. Although, predictive accuracy for ML models declined [pooled AUROC 0.803, 95% CI: 0.756-0.849] during external validation, this still exceeded the TR models external validation estimate [pooled AUROC 0.770, 0.702-0.801].
Conclusion: Machine learning shows promise for dynamic risk stratification in cardiogenic shock, yet its current predictive advantage over traditional models is modest. Due to significant data heterogeneity and performance decay in external validation, the clinical utility of these tools depends on future prospective, multi-center studies that prioritize standardized data quality and model interpretability.
Code of conduct/disclaimer available in General Terms & Conditions
{{ help_message }}
{{filter}}