The Challenge
Hospitals and healthcare organizations can often face unexpected surges in patient arrivals, leaving units like the ER unprepared and understaffed. To address this, our client sought to forecast patient census, i.e., the expected number of patients arriving at specific units up to N hours into the future. Accurate forecasting could enable proactive staffing and resource planning, cutting through inefficiencies to focus on patient care and satisfaction.
However, the organization lacked deep in-house AI expertise and faced operational, process, and cultural nuances that made integrating machine learning solutions much more difficult.
The client faced sudden patient surges that left units understaffed.
They needed accurate patient census forecasting to make their planning more proactive.
Limited AI expertise and internal nuances hindered their machine learning initiatives.
The Solution
Strategic Planning and Prototyping
We began our journey using the Xyonix Pathfinder process, which allowed us to build out a detailed understanding of their existing system and the client’s desired outcomes. We then collaborated closely with the organization to develop a robust and explainable AI forecasting solution. Our approach included:
problem formulation
Establishing a clear problem definition was crucial to ensuring alignment across their technical and non-technical teams. We set up baseline methods, starting with simple moving averages and ARIMA models for classical time series analysis. We began with univariate approaches, focusing on single variables like historical census data, before advancing to more complex multivariate models incorporating additional features.
Model Development and explainability
After initial classical approaches, we iteratively progressed to more advanced techniques, including tree-based models (e.g., Random Forest, XGBoost) and ultimately into deep learning methods, which achieved the highest efficacy.
To build trust and transparency, we implemented explainability techniques like variable holdout analysis and SHAP to identify and communicate the impact of individual variables on predictions.
The efficacy was assessed using intuitive two-dimensional plots showing error rates on the y-axis and time into the future on the x-axis. These granular plots displayed the deviation between predicted and actual census numbers, providing stakeholders with an intuitive understanding of forecast performance over time.
product design guidance
We provided strategic guidance on several product design aspects, including:
How to effectively communicate forecast efficacy in their product.
Balancing generalized models for scalability with unit- or hospital-specific models for accuracy, and explaining the trade-offs between these approaches.
Designing intuitive interfaces and workflows that could be embraced by frontline users like hospital staff.
data analysis and preparation
We worked with the client’s team to analyze and prepare their datasets. This process involved fusing multiple disparate sources into a cohesive dataset. This step was essential to capturing the nuances of patient flow across units while still adhering to strict HIPAA compliance standards.
Organizational Education and Integration
Leveraging a player-coach model, we educated their junior data science team on machine learning practices and forecasting methodologies.
On-site collaboration allowed us to deeply understand operational, process, and cultural nuances, which in turn ensured alignment between the solution and real-world needs.
Prototype Development and Feedback Loop
We developed a working prototype, enabling the organization to engage their trusted hospital clients. This feedback loop allowed the solution to evolve based on real-world insights.
The Results
Delivered significantly improved patient arrival forecasts, accurately predicting up to five days in advance.
Incorporated explainability techniques, providing transparent and easily interpretable insights for both internal teams and external customers.
Enabled client feedback through a prototype that empowered iterative product improvements.
Demonstrated AI’s feasibility and value to senior management, positioning the organization as an innovator in AI-based patient management.
Fostered technological transformation through hands-on collaboration, equipping the team with foundational AI knowledge for any future initiatives.
Have an AI project? Reach out to us.
More on our Time Series Forecasting Solutions
