JetPatient PREDICT

01Overview

PREDICT is a probabilistic ranking model that scores enrolled health-plan members by their likelihood of undergoing a defined high-cost elective surgical procedure within a forward window of 6 to 18 months. The model is designed to feed two downstream workflows on the JetPatient platform:

• The Guardian employer console, where benefits teams use ranked cohorts to time outreach, plan-design changes, and Centers-of-Excellence steerage offers.
• The PREDICT pillar surface inside the Clinic Portal, where contracted clinics receive aggregate, de-identified demand signals for capacity planning.

The model output is a per-member ranked score and a procedure-category prediction. PREDICT does not output a diagnosis, a treatment recommendation, or a coverage determination, and it is not a clinical decision support tool under FDA-defined CDS exclusions. It is a population-health prioritisation tool that requires a human reviewer in every downstream workflow.

Intended use

• Self-funded employer benefits teams (1,000+ lives) using JetPatient Guardian to design steerage offers.
• Benefits brokers and consultants using the JetPatient Broker Co-Pilot under a signed BAA.
• TPAs embedding Guardian outputs into existing case-management workflows.
• Contracted JetPatient clinics receiving aggregate, de-identified demand signals.

Out-of-scope use

• Direct-to-patient clinical decision support, diagnosis, or treatment recommendation.
• Underwriting, premium-setting, or eligibility decisions.
• Coverage determination, prior authorisation, or claims adjudication.
• Standalone deployment without a documented human-review step.
• Use on populations outside the model’s validated cohort definition (see §4).

02Model details

03Intended users

Three named roles are authorised to consume PREDICT outputs. Each role has a documented permission scope inside the Guardian console and the Clinic Portal.

Benefits team (employer or TPA)

Reads ranked cohorts at the de-identified group level. Cannot expand to member-level identifiers without an explicit BAA-scoped data-sharing agreement and a named legal basis.

Care coordinator or clinical nurse navigator

Reads member-level ranked lists for the explicit purpose of outbound coordination, only on members who have given consent at enrolment to be contacted for care navigation. All views are logged.

JetPatient network clinic operations lead

Reads aggregate, de-identified, cohort-level demand signals only. Receives no member-level information.

04Training data (v1.0 target)

Cohort definition

Adults 18–64 enrolled in a self-funded employer health plan for at least eighteen consecutive months. The training window spans thirty-six to sixty months of historical claims and eligibility, with a six- to eighteen-month forward outcome window for label assignment.

Procedure categories (v1.0 launch set)

Inclusion / exclusion

• Include: continuous enrolment ≥ 18 months; age 18–64; US residence (resident-state on file).
• Exclude: oncology pathways (separate model); pregnancy-related procedures (separate model); members enrolled fewer than 18 months at index; members lacking pharmacy benefit.
• Exclude: any member who has filed a written opt-out request through the Guardian member portal.

Data sources (v1.0)

Licensed, de-identified payer claims aggregates contracted under data-use agreements with named payer partners. Optional employer-supplied data: eligibility, voluntary biometric-screening results, and HRA responses, all transmitted under signed BAAs.

Representativeness

Training cohort balance will be reported across age decile, biological sex, US Census region, urban/suburban/rural designation, plan-funding type, and industry NAICS sector. Cohort skew vs. the US commercially insured population will be reported with a published Section 6 fairness audit.

05Performance (v1.0 target spec)

All numbers below are the target operating characteristics of PREDICT v1.0. Production deployment is blocked until validation on a real, held-out, licensed-payer cohort reproduces these characteristics within stated tolerances.

Primary operating point

Per-category performance

Per-category precision, recall, AUROC, and lead-time will be reported individually for every procedure in the §4 launch set. No category will ship with performance more than ten percentage points below the aggregate target.

06Fairness and subgroup audit

PREDICT will be audited quarterly across the following protected and quasi-protected dimensions: age band (18–34, 35–49, 50–64), biological sex, race/ethnicity where reliably captured at the payer level, US Census region, plan-funding type, employer-industry sector, and geographic urban/suburban/rural designation.

Fairness criteria

• Equality of opportunity (TPR parity across major subgroups) — target ≥ 0.90 ratio of worst-to-best subgroup TPR.
• Calibration parity — per-subgroup Brier score within 0.02 of aggregate Brier.
• Demographic parity ratio — reported, not constrained, because the underlying base rate of qualifying surgical events varies meaningfully across age and sex.
• Disparate-impact ratio (DIR) — reported in every quarterly audit; DIR < 0.80 triggers a documented mitigation review before the next release.

All fairness audits are published to the JetPatient Drift, Fairness & Audit Console. The console is the single source of truth for production deployment status and is accessible to every employer, TPA, and broker customer under their BAA.

07Limitations

• PREDICT is trained on historical claims patterns. Step-changes in plan design (formulary, network, deductible, COE program launch) introduce distributional drift; performance during the first ninety days after a major plan change is not warranted.
• Performance is lower for procedures with low base rates inside an employer cohort. Employer segments with fewer than 1,000 covered members will see wider confidence intervals and are flagged as such in the Guardian console.
• The model is not validated outside US-domiciled commercially insured populations. Medicare, Medicaid, and Tricare cohorts are out of scope for v1.0.
• The model is not validated on populations with predominantly cash-pay or international care-seeking patterns. International cross-border care signals are not in the v1.0 feature set.

08Monitoring, maintenance, and rollback

Drift monitoring

PREDICT runs a weekly drift audit covering input-feature distribution (population stability index, Kolmogorov–Smirnov), output-score distribution, and observed-vs-expected outcome rate on the prior month’s scored cohort. Drift dashboards are visible to the model owner, the clinical advisory board, and customer compliance officers.

Retraining triggers

• Scheduled: quarterly retrain on rolling-window licensed-payer data.
• Drift-triggered: PSI > 0.25 on any feature, or month-over-month output-score KS > 0.15.
• Fairness-triggered: any subgroup TPR drop > 0.05 quarter-over-quarter, or DIR < 0.80.
• Outcome-triggered: observed-vs-expected event rate outside the 95% prediction interval for two consecutive months.

Rollback

Every PREDICT release is reversible. The Drift, Fairness & Audit Console exposes a one-click rollback for the model owner, with audit-logged justification required. The prior production version is retained for at least four prior generations.

09Privacy, compliance, and governance

Data handling

• All training data is processed under Safe Harbor de-identification or a HIPAA-compliant Limited Data Set, governed by an executed Business Associate Agreement with the data source.
• Production scoring uses the minimum-necessary feature set per signed BAA with each employer or TPA customer.
• PREDICT does not retain member-level features after scoring beyond the contracted retention window. Default retention: 24 months, configurable per customer.
• Encryption: AES-256 at rest under customer-isolated KMS keys; TLS 1.3 in transit; no plaintext at any intermediate hop.

Governance

• Model approval requires sign-off from the JetPatient clinical advisory board (named clinical lead, biostatistics lead, and patient-advocacy representative).
• Every production release is logged in the model registry with: training-data hash, code commit, hyperparameter configuration, full evaluation report, and fairness audit.
• Incidents (model drift, fairness violation, customer-reported error) are tracked in a public post-incident review queue with a 14-day publication SLA.

10Version log

↓ Download Model Card v2.4.1 (PDF) Back to Trust & Methods hub