The Clinical Prediction Engine (CPE) is a project sponsored by The Zodiac Group in partnership with Humber Polytechnic. The CPE was designed to cater to at-risk patients who prefer or are advised to stay at home, formally described as remote patients.

Remote patients have a higher risk of not getting the immediate attention needed during an emergency, and to know their potential for an emergency using risk scores switches emergency aid from being reactive to a proactive response, increasing monitoring and response alertness for patients with increasing risks.

Using our Vitagent application as the springboard for this project, the CPE aimed to implement the proactive emergency response aspect for remote at-risk patients by using deep learning models to predict a patient's health risk using their historical and current vital signs.

Complex Deep-Learning Neural Network: Uses the Temporal Fusion Transformer (TFT), a customized deep-learning network using Transformers and gated residual networks to predict the patient's health risk using their current vital signs and their recent medical history.

Low-cost Lightweight Machine Learning Algorithm: Using results attained for the Vitagent project, the XGBoost Algorithm was also implemented as a baseline model for comparison, but using updated training data from preprocessing improvements.

Explainable AI (XAI): Uses the SHAPley explainability package for XGBoost to provide feature-based details to for every prediction. TFT on the other hand, produces values using charts generated from each of its components - the variable weights from the Variable Selection Network (VSN) and the attention weights from the transformer itself

Informative Risk Scoring Framework: The RSTF provides a proper channel for interpreting a select model's output into suggestive and interpretable information presented back to the user. The presentation includes a Risk Score for a granular representation of a patient's chance to have an emergency from 1-10, a Risk Tier as the tier updated from the risk score, and the a degree of error to highlight the model's confidence in that prediction.

As part of our instructor's advise before officially embarking on our capstone sponsor's AI project, whom we consider as the main project stakeholder, we were instructed to complete a project with a reduced scope based on our sponsor's requirements. The goal was to become familiar with the stakeholder's needs that should be addressed. To that end, we worked up an alpha version using the previous phase of the upcoming project which used XGBoost as the baseline model — also outlined on our project roadmap. As such, Project Vitagentcame into fruition.

Vitagent transitions patient monitoring from reactive alerts to proactive forecasting leveraging XGBoost's gradient-boosting algorithm. By analyzing trends in vital signs, Vitagent assesses health risks independently of underlying conditions, providing clinicians with actionable, data-driven foresight outside of the traditional black-box operations of most predictive models. Using XGBoost to read feature changes over time, the patient's health risk is predicted using gradiented tree-based algorithm, achieving at least 82% overall accuracy by project completion. The prototype also uses a serverless architecture through AWS Lambda, allowing users to do in-demand usage and keep application hosting costs minimal.

Trend-Based Analysis: Emulates EHR workflows by assessing risks based on changes between medical records over a rolling time window.

Two-Layer Intelligence: A 2-layered system designed to "Predict and Suggest," identifying Low-Risk or High-Risk states as data density increases.

Shapley Explainable AI (XAI): Every suggestion is backed by a SHAP explanation, detailing the specific vital signs that influenced the model’s assessment.

Serverless Application Stack: The prototype is hosted using an AWS serverless pipeline with AWS Cloudfront, AWS API Gateways and 2 Lambdas hosting the two-layered intelligence system

Project Vitagent was not without flaws. One of the main flaws Kevin discovered (only after project completion) was that if a patient's vitals remained constant (always the same low / high values) over a period of time, the slope feature between changes approaches zero, and causes the predictions to become biased to low risk even with consistently high-risk readings. This was vital information we needed to learn before the actual project, and due to this finding, we were able to avoid this pitfall in the actual project itself.

Another was that deploying the trained xgboost model with its prerequisite packages needed over 200MB volume size - which was way past a regular Lambda's allowed limit. For that, the team learned how deploy ECR images into lambda to cover the serverless aspect, which while imperfect due to the slow cold start, still got the job done.