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

Technical Summary: External Validation of Machine Learning Models for Traumatic Brain Injury

Problem Statement
The deployment of Machine Learning (ML) and Deep Learning (DL) models for Traumatic Brain Injury (TBI) prediction faces a critical trade-off between predictive performance and operational feasibility. While complex "multiparameter" models integrating diverse data sources (clinical variables, biomarkers, radiomics) theoretically offer superior information gain, they require extensive computational resources, fragmented data integration, and significant energy consumption. Conversely, "pauci-parameter" models rely on limited, routinely available variables but may lack the granularity of complex systems. A knowledge gap exists regarding whether the marginal gains in predictive accuracy of complex models justify their high computational costs, carbon footprints, and deployment barriers, particularly in resource-limited settings or for real-time clinical decision support.

Methodology
This study conducted a retrospective, single-center external validation at Grenoble University Hospital (a Level 1 trauma center) involving 534 adult TBI patients (GCS <9 and/or intracranial injuries). The study compared seven distinct ML models across four categories of outcomes: neurocritical care intensity (TILSUM and SIBICC classifications), mortality (7-day, 30-day, 6-month), and functional outcome (Extended Glasgow Outcome Scale).

The models were categorized as follows:

1. Pauci-Parameter Models (Resource-Efficient):
   • PREHOSP: Utilized 15 routine prehospital variables (e.g., GCS, vital signs, pupil reaction) using Light Gradient Boosting.
   • CT-TIQUA: Analyzed admission CT scans (within 60 minutes of arrival) using a Convolutional Neural Network (CNN) pipeline involving bone removal, topographic registration, segmentation, and quantification of seven injury classes (e.g., hematoma, edema).
2. Multiparameter Models (Data-Intensive):
   • Five models integrating broader datasets: MULTI (PREHOSP + CT-TIQUA features), PREHOSP-X (23 prehospital variables), RESUS-X (18 resuscitation room variables), MULTI-PRE (PREHOSP-X + CT-TIQUA), and MULTI-RESUS (RESUS-X + CT-TIQUA).

Evaluation Metrics
The study employed a multi-dimensional evaluation framework:

• Predictive Performance: Primary outcome was the Positive Likelihood Ratio (LR+) with 95% confidence intervals. Secondary metrics included AUC-ROC, sensitivity, specificity, Brier score, and Net Clinical Benefit (at a 10% decision threshold).
• Efficiency: Computation time for training and single-case inference was measured in seconds.
• Environmental Impact: Carbon footprint (kg CO₂eq) was calculated from development to application using the CodeCarbon library, accounting for energy consumption (kWh) and local carbon intensity.

Key Results

• Predictive Performance: Multiparameter models demonstrated superior statistical performance in specific areas, particularly for mortality prediction. The best-performing multiparameter model (MULTI-PRE) achieved the highest LR+ for 6-month mortality (41.83) compared to the pauci-parameter PREHOSP model (14.19). However, for neurocritical care intensity predictions (TILSUM, SIBICC), multiparameter models did not show systematic superiority, with LR+ values generally remaining below 10 for all models.
• Clinical Utility: Despite differences in statistical metrics, the Net Clinical Benefit was nearly identical across models. At a 10% decision threshold, the complex MULTI-PRE model yielded a benefit of 0.06 (6 additional correct decisions per 100 patients), while the simple PREHOSP model yielded 0.05.
• Efficiency and Carbon Footprint: The efficiency gap was substantial. The deep learning model (CT-TIQUA) required approximately 100 times longer for inference (342 seconds vs. 0.0034 seconds for Gradient Boosting) and generated 10 times higher carbon emissions per prediction compared to simpler models. Training the CNN model generated 377.7 g CO₂eq, equivalent to six clinical CT scans in the French energy mix, whereas the Gradient Boosting model required only 1.3 g CO₂eq.
• Generalizability: Models relying solely on prehospital data (PREHOSP) demonstrated robust performance without requiring access to complex hospital electronic health record (EHR) ecosystems or immediate CT imaging, suggesting higher feasibility for deployment in diverse settings.

Significance and Claims
The paper argues that computational complexity and resource intensity must be weighed equally against predictive performance in clinical AI deployment decisions. The study challenges the assumption that "more data equals better outcomes," demonstrating that for TBI pathway prediction, simpler pauci-parameter models can achieve comparable net clinical benefit to complex multiparameter models while drastically reducing environmental impact and deployment barriers.

The authors claim this is one of the first validation studies to systematically compare the carbon footprint of clinical prediction models against their performance. They conclude that for sustainable digital health implementation—especially in low-resource settings or where EHR integration is fragmented—simpler models offer a viable, efficient alternative that enables broader access without sacrificing clinical utility. The study recommends reserving complex, resource-intensive models for specialized high-resource centers where infrastructure supports them, while advocating for the validation and deployment of simple, prehospital-based models in emergency medical services and regional hospitals.