HealthTech Startup Transforms Reactive Discharge Tracking into Real-Time Treatment Risk Prediction with Claude and Amazon SageMaker
At a glance
Behavioral health treatment providers can rarely tell, in the moment, which patients are at rising risk of leaving care against medical advice. By the time a care team notices, the patient has already disengaged, and the relapse and readmission risk that follows is a well documented pattern across the addiction and mental health treatment field nationally.
New Math Data built this HealthTech startup a HIPAA-ready machine learning pipeline on AWS that scores active patients for AMA risk using clinical assessments and engagement data already inside the platform. Care teams can now flag at-risk patients before they leave treatment, not only after.
The company moves from a retrospective outcomes-reporting platform to one with embedded predictive intervention, a capability its provider and payer network can point to directly as it competes for value-based, outcomes-accountable contracts nationwide.
Industry
Use Case
Machine Learning, Risk Assessment & Management, MLOps & Platform Modernization
Solution implemented
- XGBoost model deployed on Amazon SageMaker predicts treatment completion versus AMA discharge risk.
- Asynchronous API on ECS Fargate behind a private ALB serves predictions via callback or polling.
- SageMaker Model Registry governs versioning, F1-based approval thresholds, and safe rollback.
- Terraform-defined infrastructure automatically redeploys on container or pipeline configuration changes.
- SQLAlchemy and Polars power modular feature engineering from clinical, survey, and placement data.
- An earlier Claude on Amazon Bedrock assistant enables natural-language patient data analysis for staff.
The value equation
- Care teams can flag AMA risk before a patient leaves treatment, not after.
- The company moves from retrospective reporting to real-time predictive clinical intervention.
- HIPAA-ready production pipeline replaces manual review with automated, explainable risk scoring.
- Model registry and Terraform-based deployment let the team retrain and redeploy without vendor lock-in.
- A follow-on services engagement positions NMD as ongoing partner for cohort-specific model expansion.
Company Snapshot
A Philadelphia-based behavioral health technology startup providing outcome-tracking and population health management to treatment providers, managed care organizations, and self-funded payers nationwide.
Location
United States
Customer Situation
This company’s outcome-tracking platforms track behavioral health outcomes for providers, managed care organizations, and self-funded payers nationwide. Clinical teams could see how patients were doing, but had no forward-looking signal for who was at risk of leaving treatment against medical advice, a disengagement event tied to higher relapse and readmission risk.
Care coordinators were left to react after a patient had already left rather than intervene while there was still a chance to change the outcome, even as payers pushed providers toward value-based, outcomes-accountable contracts.
This gap mirrors a documented pattern across behavioral health and substance use treatment nationally, where discharge against medical advice remains a persistent driver of relapse and readmission.
NMD Solution
NMD reviewed the company’s clinical and engagement data and found that treatment completion could be modeled directly from placement, assessment, and survey records already inside the platform, without new data collection. The solution required a HIPAA-ready, asynchronous machine learning architecture: an XGBoost model on Amazon SageMaker, an ECS Fargate API behind a private Application Load Balancer inside a secured VPC, and a governed model registry for versioning, approval thresholds, and rollback, all defined in Terraform for repeatable, environment-consistent deployment.
What We Delivered
NMD delivered a production-ready machine learning pipeline that scores active patients for AMA risk using clinical assessments, engagement history, and placement data, with Shapley-based feature explanations clinicians can interpret directly against clinical expectations. The company’s team is now integrating the service into its HIPAA-compliant production environment, with on-demand support from New Math Data during rollout. Documentation, runbooks, and recorded knowledge transfer sessions equip the team to retrain, extend, and operate the system independently going forward.
Ready to Transform Your HealthTech Business?
See how New Math Data can transform your HealthTech business with AWS-powered innovation.