Machine Learning vs Humans: Identifying Trauma Patients in Need of Haemorrhage Control Resuscitation

Timely identification of haemorrhage control resuscitation (HCR) needs in trauma patients is challenging due to inconsistent recognition and decision-making, which delay treatment and harm outcomes. Traditional clinical tools like scores and flowcharts are easy to use but often lack predictive accuracy and are poorly integrated into clinical practice. 

Machine learning (ML) offers a promising alternative by providing automated, real-time decision support, even with incomplete data. While many ML models exist, few have undergone external or real-life validation. 

The ShockMatrix study aimed to address this by developing and testing an ML algorithm, supported by a smartphone app, to predict HCR needs using routine prehospital data. The study compared the ML model’s accuracy with clinician judgment, adhering to STARD 2015 and DECIDE-AI guidelines, and aimed to evaluate both predictive performance and real-world integration into trauma care.

This study was conducted across eight level-1 trauma centres. Trauma clinicians entered nine predefined predictor variables into a smartphone app and gave their prediction regarding the need for HCR. These inputs matched those used by the machine learning model. The primary outcome was the need for HCR, defined as transfusion in the resuscitation room, transfusion of more than four red blood cell units within 6 hours, a haemorrhage control procedure within 6 hours, or death from haemorrhage within 24 hours. The performance of clinicians and the machine learning model was evaluated using sensitivity, specificity, likelihood ratios, net clinical benefit, and inter-rater agreement via Cohen’s kappa coefficient.

1,292 of 5,550 screened trauma patients were included in the study, with 13% (170 patients) requiring HCR. Human predictions had a positive likelihood ratio (PLR) of 3.74 and a negative likelihood ratio (NLR) of 0.36, while the machine learning model had a PLR of 4.01 and an NLR of 0.35. When combined, human and machine predictions achieved 83% sensitivity and 73% specificity. The agreement between human and machine predictions was moderate at 0.51.

The ShockMatrix study demonstrated that an ML model using only nine routine prehospital variables could predict the need for HCR with accuracy comparable to experienced trauma clinicians. The moderate agreement between human and ML predictions suggests their combined use may improve decision-making.

This study fills a key gap by prospectively validating an ML model in real-time and benchmarking it against human predictions in acute trauma settings. Data collection was efficient via a smartphone app, and the model produced actionable outputs, potentially reducing delays in blood product administration—an issue highlighted in recent trials like CRYOSTAT-2.

By supporting early decision-making, the ML tool may reduce cognitive load, enhance guideline adherence, and improve patient safety. While traditional scores like the Shock Index and ABC score are easier to use, they often lack accuracy, rely on outdated cohorts, and fail to reflect current clinical needs. In contrast, ML models can process continuous data and adapt to current standards.

The ShockMatrix algorithm performs on par with other advanced models but emphasises real-time, actionable use. Unlike many retrospective models requiring data not available prehospital (e.g., AIS scores), this model is optimised for early trauma care.

The study shows ML doesn’t need to outperform clinicians to be valuable; its strength lies in complementing human judgment. A randomised trial planned for 2025 across 16 French dispatch centres will evaluate this integration, including models for HCR, neurosurgical intervention, and intracranial pressure monitoring, along with cost-effectiveness analysis.
 

Source: The Lancet 

Image Credit: iStock