Approach

To improve admission forecasting, the CIC developed an AI-driven prototype that analyzes anonymized hospital data to predict ED patient admissions. With the goal of reducing ED boarding, the system assigns real-time admission scores and presents them on an interactive dashboard, making it easy for hospital staff to prioritize patients and resources. Designed to integrate smoothly with hospital workflows, it ensures that resourcing managers, pharmacy technicians, and clinicians can easily access and act on predictions. Built on AWS services like Amazon SageMaker and AWS Lambda, the solution is scalable, efficient, and secure. By automating admission predictions, it helps hospital staff make faster, data-driven decisions, leading to better patient flow and resource management.

Supporting Artifacts

Through an iterative process of reviewing previous literature, model experimentation and testing, and data refinement, we ended up with two candidate classifiers best suited for the task – a Random Forest classifier (RF) and an XGBoost classifier (XGB). During testing, we compared performance using different metrics in an 8-fold cross-validation evaluation schema. In each run of cross-validation, a unique training set and test set are chosen, a Grid-search is run choosing optimal hyper-parameters for each model and metrics for evaluation are averaged across all folds. Below are the results in terms of the metric AUC (area under the curve) which gives us a general idea of the effectiveness of the classifiers.

Further, looking at the accuracy and class accuracies, we found that these models were good at predicting non-admissions. For practical purposes, it is better for the model to not turn away a patient who may actually be admitted and in need of a bed, thus erring on the side of accurately predicting non-admissions with a few false positive alerts greatly outweighs a model with a high false negative rate. This AI model aims to be useful for ED staff by alerting them about patients who are likely to be admitted while confidently predicting those who would not likely be admitted.

Thus, our solution enabled a real-time inferencing model along with a dashboard. The full architecture of the solution can be seen below.

Architecture Diagram

Technical Details

1. The training pipeline is separated on a different platform (Amazon SageMaker Notebook) to allow for users to experiment and view model results.
2. The user can upload labeled training data (as .csv) to the Training data s3 bucket. This bucket will store the labeled training data as well as the default inference script and requirements files (which will be used to create the Amazon SageMaker model endpoint later)
3. The user can interact with Amazon SageMaker Notebooks to experiment and view model results. Accordingly, the user can launch an Amazon SageMaker training job to save and store a final model.
4. The training job will package the model along with the required inference script and requirements files in a “model.tar.gz” as part of the training output 
5. The user can upload the “model.tar.gz” file to the endpoint s3 bucket which will process an AWS Lambda to upload the model as a Amazon SageMaker endpoint
6. The user can upload new patient data files (as .csv) to the inference bucket using SFTP. This will trigger an AWS Lambda to automatically infer the model and store the output to Amazon Aurora.
7. The front end will refresh based on the data in the Amazon Aurora table through an AWS Lambda trigger. The patient dashboard will showcase the model output along with the time that the front end was most recently updated.

User Interface

This real-time interface displays key patient details, including facility ID, registration time, and model generated admission scores, enabling healthcare providers to assess cases quickly.

With built-in filtering and sorting options, clinicians can refine the data based on facility and urgency, allowing for a more focused view of critical cases.

Link to solution on GitHub: https://github.com/UBC-CIC/Hospital-Admissions-Forecasting

Infographic

Acknowledgements

Project Sponsors: Halton Healthcare

Student team: Developers: Rohit Murali and Khushi Narang. Project Assistance by Amy Cao.