This article presents a commentary on the role of machine learning (ML) and predictive models in the management of traumatic haemorrhage. The use of ML in predicting patient outcomes can guide clinicians in making treatment decisions for haemorrhaging trauma patients. However, the current approach of binary outcome prediction based on previous clinical decisions has its limitations.
The decision to transfuse an injured patient relies on the clinician’s assessment of the patient’s need for blood products. Transfusion is often time-critical and can be challenging due to uncertainty in decision-making. Various scoring systems exist to aid clinicians in making these decisions, but their clinical uptake has been limited.
One limitation is the arbitrary criteria for denoting massive transfusion, which is problematic as patients predicted to require slightly more or fewer units may have similar physiology and outcomes. Predicting nominal categories, such as the likelihood of consuming 20, 30, or 50 units of blood, can provide better granularity and inform triaging decisions to facilities with sufficient blood stocks.
Another limitation is the reliance on historic clinical choices embedded within ML training data. Predicting the true patient state, i.e., the underlying requirement for transfusion, offers more value to decision-makers and reduces the risk of propagating under- or over-transfusion practices. This approach is less affected by biases and trends in practice.
Furthermore, these limitations may contribute to survivorship bias, where ML models only learn from patients who meet a certain threshold for transfusion. This bias can skew the accuracy of predictions and hinder the overall effectiveness of predictive models.
In conclusion, while ML and predictive models have the potential to improve decision-making in traumatic haemorrhage care, the current approach of binary outcome prediction and reliance on previous clinical decisions has its limitations. Moving towards predicting nominal categories and the true patient state can provide better granularity and reduce biases in transfusion practices. Further research is needed to refine and optimize predictive models for traumatic haemorrhage management.
Sources:
– “Artificial intelligence and machine learning for hemorrhagic trauma care” by Peng et al. (source article)
– Definitions:
– Traumatic haemorrhage: severe bleeding caused by trauma or injury
– Machine learning (ML): the use of algorithms and statistical models to enable computer systems to learn and improve from experience without explicit programming.
– Binary outcomes: outcomes that are expressed in two categories, such as yes/no or true/false
– Survivorship bias: the bias that arises when analyzing only the successful or surviving cases and ignoring the failure or non-surviving cases.
