August 30
@
12:00 PM
–
1:00 PM
London Business School
Faculty Headshots – 25.05.22
Jean Pauphilet
Photo © Sheila Burnett
With the growing availability of patient-level data and advanced data analytics tools, hospitals are increasingly working to develop prediction models that leverage this rich data to improve operational efficiency. However, analyzing raw patient-level data and deploying real-time analytics that rely on them brings another set of challenges for hospital systems. In this context, evaluating the operational value of patient-data analytics is crucial for hospital managers.
In this talk, we investigate two uses cases.
Patient-level discharge prediction. We collaborated with a large hospital network in the US, developed models to predict short- and medium-term outcomes for all inpatients across their 7 hospitals, and deployed the models as part of a user-friendly dashboard and a color-coded alert system to communicate these patient-level predictions. Since its deployment, over 200 doctors, nurses, and case managers have been using the tool in their daily patient review process. We find that doctors start the administrative discharge process earlier, leading to a significant reduction in the average length of stay (0.63 days per patient). This talk will emphasize the challenges of deploying and of identifying the main value added of such tools.
Emergency Department (ED) occupancy prediction. For many resource allocation decisions, visibility on overall occupancy levels are sufficient for an informed decision-making. Yet, there is a growing interest in using patient-level data to predict discharge of each patient and aggregate them into a unit-level prediction for the entire ED. We ask the following questions: Are such bottom-up approaches significantly more powerful? Are patient-level models the most efficient way to use granular patient-level information? We find that time-series models, despite their simplicity and relatively low complexity, significantly benefit from the inclusion of summary information about current ED patients, leading to a 15-30% reduction in Mean Absolute Error (MAE). Bottom-up approaches, offering dual-level (patient and ED) predictions, not only improve interpretability, but also significantly improve occupancy prediction accuracy. However, to achieve these gains, we show that bottom-up pipelines require an additional calibration step (and careful re-calibration over time). We also compare with deep set neural networks that predict occupancy directly from the concatenation of all patient-level information available.