Combining Cognitive and Machine Learning Models to Mine CPR Training Histories for Personalized Predictions
Florian Sense, Michael Krusmark, Joshua Fiechter, Michael G. Collins, Lauren Sanderson, Joshua Onia, Tiffany Jastrzembski
Jul 01, 2021 15:05 UTC+2
—
Session D1
—
Zoom link
Keywords: Predictive modeling, Cognitive model, Machine learning, Cardiopulmonary resuscitation, Learning
Abstract:
Cardiopulmonary resuscitation (CPR) is a foundational life-saving skill for which medical personnel are expected to be proficient. Frequent refresher training is needed to prevent the involved skills from decaying. Regular low-dose, high-frequency training for staff at fixed intervals has proven successful at maintaining CPR competence but does not take into account inherent performance variability across learners. Tailoring refreshers to an individual’s past performance, would minimize personnel being trained too (in)frequently and would ensure faster knowledge acquisition for new learners. To maximize the benefits of individualized schedules, learning needs gleaned from past training history need to be identified. Predictive analytics modeling tools can be used to forecast when a given learner should return for refresher training to maintain a predefined performance standard. A recent field study conducted among nursing students showed that a cognitive model-based approach was able to successfully trace the knowledge acquisition and decay of learners and prescriptively devise personalized training regimes that outperformed fixed schedules with regards to both training efficiency and learners’ performance. Here, we report a post-hoc analysis of the collected data to investigate whether an alternative modeling approach, blending cognitive modeling and machine learning, could have resulted in even higher quality predictions. Our results reveal modest improvements in predictive accuracy for ensemble models, in which machine learning models predict the prediction errors (i.e., residuals) of the standalone cognitive model. These promising findings reveal strong applied utility for future use in domains where sustained proficiency is a requirement.