An Assessment of Machine Learning Simulation in Clinical Decision Making

Though machine learning algorithms have become ubiquitous in many fields, their study has just recently begun in medicine. This has sparked a desire to explore novel approaches to improving patient outcomes.

One common application of machine learning models for clinical decisions is making prognoses about whether a patient will experience an adverse outcome. The provider, who is directed by the algorithm's predictions, is a vital aspect in the implementation of any machine learning model in healthcare.

Value assessments of machine learning algorithms, which consider patient outcomes and cost savings, may be complicated by the presence of the provider, who introduces several key operational concerns. These concerns include limited availability of providers, prediction timing, provider fees, intervention success rates, and cost savings from pathway interventions.

The majority of predictive algorithms needing pathway implementations are assessed with predictive performance metrics, such as the c statistic. Although these metrics and approaches help understand predictive performance, they are not clinically interpretable because they ignore resource constraints and the effect of the algorithm's prediction on the patient's health. They do not consider the clinicians' schedules or the limits on how many patients they can physically accommodate on a given day.

Clinicians and hospital administration can evaluate the efficacy of a predictive model in a dynamic, resource-limited context using machine learning simulation. The approach entails monitoring patients from the time they leave the operating room until they are discharged, determining the risk of each patient using the predictive model, and selecting the highest risk patients for intervention based on the provider's availability and the maximum number of patients the provider can handle within a day.

Conclusion

The simulation-based evaluation model is designed for clinicians and hospital management to use a machine learning model in real-time to manage the allocation of limited resources. Finally, the technique enables clinicians and administrators to assess the value of any prediction model in terms of patient outcomes and costs in a more clinically useful manner.