Is simple better? Comparing Computational Cost and Carbon Impact of Machine Learning Models for Traumatic Brain Injury Prediction; A Case Study for Sustainable Digital Health Implementation

Authors

Abstract

BackgroundMachine learning (ML) models for traumatic brain injury (TBI) prediction increasingly demand extensive data, computational resources, and energy consumption, yet simpler models may offer comparable clinical benefit with lower barriers to deployment. This study compares predictive performance, computational efficiency, carbon footprint, and real-world feasibility of resource-efficient ("pauci-parameter") versus data-intensive ("multiparameter") ML models for predicting TBI patient care pathways and outcomes. MethodsExternal validation study in a level 1 trauma center (n=534 adult TBI patients with GCS<9 and/or intracranial injuries). Seven models tested: two pauci-parameter models using only routine prehospital variables (PREHOSP, 15 variables) or CT image analysis (CT-TIQUA), and five multiparameter models integrating clinical and imaging data. Primary outcome: positive likelihood ratio for predicting neurocritical care intensity, mortality (7/30-day, 6-month), and functional outcome (Glasgow Outcome Scale Extended). Secondary outcomes: computation time, carbon footprint, clinical implementability. ResultsMultiparameter models showed superior performance but did not consistently translate to better clinical utility. PREHOSP (pauci-parameter) showed comparable performance to complex models for most outcomes. The best-performing multiparameter model (MULTI-PRE) required 100-fold longer inference time and 10-fold higher carbon emissions per prediction versus simple models, while net clinical benefit was nearly identical (0.06 vs 0.05). Models using only prehospital data demonstrated greater generalizability and lower deployment barriers. Interpretation Computational complexity and resource intensity should factor equally with predictive performance in clinical AI deployment decisions. For sustainable digital health implementation--especially in resource-limited settings--simpler models with comparable clinical benefit may enable broader access while reducing environmental and financial costs. FundingFondation Gueules Cassees, Grant 27-2023 AUTHOR SUMMARYO_ST_ABSWhy Was This Study Done?C_ST_ABSArtificial intelligence (AI) models are increasingly used in hospitals to help doctors predict which brain injury patients need intensive care and what their outcomes might be. However, many published models are very complex. They use multiple variables and require expensive computer systems to run. Hospitals often struggle to implement these complex models because they require operationalisation barriers. - Multiple data systems that dont easily communicate - Expensive computer infrastructure - Highly trained technical staff - Energy-intensive processing This raises the question whether complex models actually work better than simpler ones? If simple models work just as well, they could be used in many more hospitals especially in lower-resource countries where complex systems arent available. What Did the Researchers Do?The team compared seven different Machine and Deep Learning models for predicting brain injury patient needs and outcomes in 534 patients treated at a major trauma center in France. They tested: - Two simple models using only routine prehospital data (vital signs, Glasgow Coma Scale) - Five complex models using 40+ variables from multiple hospital computer systems plus CT scan segmentation analysis They compared the models accuracy, how long they took to run, how much electricity they used (carbon footprint), and how easy they would be to implement in different types of hospitals. What Did They Find?The simple models worked almost as well as the complex ones. For example: - Simple model (PREHOSP): Predicted mortality with reasonable accuracy - Complex model (MULTI-PRE): Predicted mortality slightly better, but required 100 times more processing time and 10 times more electricity When measured by "clinical benefit", the number of correct treatment decisions made per 100 patients, both simple and complex models performed similarly (0.05 vs 0.06 additional correct decisions). The complex models training generated as much greenhouse gas as a clinical CT scan. For hospitals committed to environmental sustainability, this matters. Why Is This Important?This study challenges the common assumption that more complex models perform better. The findings have practical implications: For well-resourced hospitals: Complex models may offer only modest benefit that may not justify their cost and complexity. For hospitals with limited budgets: Simple models may provide comparable accuracy without expensive infrastructure. For low-income countries: Simple models using prehospital data may be deployed immediately in emergency medical services, even without integrated hospital computer systems. For environmental sustainability: Simpler AI models consume far less energy, an important as hospitals worldwide commit to net-zero emissions and will implement energy intense decision support tools. What Are the Clinical Implications? The authors recommend: O_LIAsses clinical performance, operational barriers and carbon imprint when comparing simple and complex models before implementation and deployment of decision support tools C_LIO_LIReserve complex models for specialized high-resource trauma centers where infrastructure supports them C_LIO_LIValidate simple models in diverse care settings (ambulance services, regional hospitals, low-income countries) to confirm they work everywhere C_LI What Comes Next?The next step is testing the simple model in real-world settings across different countries and hospital types to prove it works reliably everywhere.

Tags

Topics