Regional hospital network · 8 hospitals
The hospital network had a 30-day readmission rate of 14.2% — above both the CMS national average and the network's own target. CMS readmission penalties were costing the network approximately $1.8M annually. The clinical leadership understood that readmissions were often preventable — patients discharged without adequate follow-up plans, at-risk patients not identified before discharge, care coordination gaps between inpatient and outpatient teams — but lacked the tooling to identify high-risk patients systematically at the point of discharge.
The network had tried to use the standard readmission risk scoring tools available in their EHR, but clinical staff did not trust them. The scores were produced by models the clinical team could not interrogate, and the recommendations that followed from them were not connected to the specific risk factors present in an individual patient's record.
Shelorve began with a data readiness assessment before any model design work. The network's clinical data — five years of admission, discharge, and transfer records, diagnosis codes, medication histories, and care coordination notes — was distributed across three systems. Shelorve built a data pipeline on AWS (Kinesis for real-time feeds, S3 for the data lake, Redshift for the analytical layer) that brought these sources together into a unified patient record before the model training began.
The readmission risk model was built on AWS SageMaker, trained on four years of historical admissions with confirmed readmission outcomes. Rather than predicting readmission probability as a single score, the model outputs a risk stratification with the contributing factors weighted by importance for each individual patient — so a clinician reviewing the discharge dashboard sees not just "high risk" but "high risk primarily due to medication adherence history and absence of outpatient follow-up appointment."
The explainability layer was built using SageMaker Clarify, and the output was designed specifically for clinical use — plain-language risk factor descriptions rather than feature importance coefficients. The model was integrated with the EHR via HL7 FHIR, so risk scores appear in the clinical workflow at the point of discharge planning without requiring clinicians to access a separate system.
30-day readmission rate fell from 14.2% to 11.5% in the first year — a 19% reduction. CMS readmission penalties fell by $340,000 in the first year of operation. Clinical adoption was high from the first month: the discharge planning team credited the explainability output as the reason they trusted and acted on the model's recommendations, where they had not trusted the previous EHR-native scoring tools. The model operates in production at all 8 hospitals with no manual intervention.
"The previous readmission tools gave us a number. This gives us a reason — and the reason is specific to the patient in front of us. That is what changed the clinical team's relationship with the model."
SageMaker · SageMaker Clarify · Kinesis · S3 · Redshift · QuickSight · API Gateway (HL7 FHIR)
The key difference was explainability. The previous tools returned a readmission probability score with no supporting rationale. Clinicians could not act on it with confidence because they could not validate whether the model was assessing the right factors for the specific patient. The Shelorve model returns patient-specific risk factors in clinical language. A discharge coordinator sees a recommendation grounded in the patient's actual record — not a black-box score.
All data is encrypted in transit and at rest. SageMaker endpoints are deployed in a VPC with no public internet exposure. The data pipeline from EHR systems to the S3 data lake uses encrypted Kinesis streams. BAAs are in place with all AWS services. The QuickSight dashboards used by clinical staff access Redshift through VPC-native connections. Shelorve provided a HIPAA compliance evidence pack covering all components at deployment.
The model includes a confidence score alongside the risk stratification. For patients with limited data — new to the network, or with sparse historical records — the confidence score is lower and the clinical dashboard surfaces this explicitly. Care coordinators treat low-confidence, high-risk scores as a reason to gather more information at discharge, rather than acting on the model output alone. The discharge planning protocol was updated by the clinical governance team to reflect these confidence-weighted workflows.
Tell us what you are trying to fix. We will tell you whether Shelorve is the right partner — and if we are not, we will tell you that too.