Large Claimant Prediction
Predictive modeling for catastrophic claim probability with machine learning for high-cost claimant identification and risk quantification.
The $1.5M Shock No One Saw Coming
Large Claims Hit Balance Sheets Like Lightning
A 1,000-life self-insured employer with a $250K stop-loss deductible can easily absorb one catastrophic claimant per year. But three simultaneous large claimants—a premature birth ($850K), a cancer diagnosis ($620K), and an accident with long-term care ($480K)— creates $1.95M in total claims, of which $750K is below the deductible and hits the balance sheet directly.
Problem: Traditional risk models don't predict these outliers. They surface AFTER someone has already incurred $150K-$200K in claims—too late to intervene, manage care trajectory, or adjust stop-loss coverage.
Without Predictive Intelligence
- ×Large claimants only identified AFTER they've incurred $100K-$200K
- ×No opportunity to steer to Centers of Excellence or high-value providers
- ×CFO gets blindsided by claims volatility at month-end close
- ×Stop-loss deductibles set reactively based on last year's experience
- ×No data to negotiate with stop-loss carriers (they have pricing power)
With Large Claimant Prediction
- Identify high-risk members 6-18 months before they cross $100K threshold
- Proactive care navigation to high-value providers (save 20-40% on episode cost)
- Dynamic stop-loss pricing with carrier (quantified risk = negotiating leverage)
- Budget forecasts include probabilistic large claim exposure (P50/P75/P90)
- CFO dashboard with real-time large claimant pipeline and financial exposure
ML Model Architecture
Three-Stage ML Pipeline
Stage 1: Multi-Modal Feature Engineering
1,847 FEATURESSynthesize signals from medical claims (diagnosis patterns, procedure trends, specialty visits), pharmacy (drug classes, adherence, dose escalations), biometrics (HbA1c trends, BMI trajectory), and social determinants (zip code SVI, commute time to specialty care). Time-series features capture acceleration (e.g., "3 cardiology visits in 90 days after zero in prior 2 years").
# Feature Categories
Medical Claims (840 features):
- Diagnosis code sequences (ICD-10 n-grams)
- Procedure frequency (CPT codes, rolling 90/180/365 day windows)
- Specialty visit patterns (oncology, cardiology, neurology)
- Inpatient utilization velocity
- ER visit frequency and acuity codes
Pharmacy (520 features):
- Drug class fills (antineoplastics, biologics, GLP-1s)
- Medication possession ratio (MPR) trends
- Dose escalation patterns (e.g., insulin dose × 3 in 6 months)
- Days supply variability
- Prior authorization denials
Biometrics (310 features):
- Lab result trends (HbA1c, creatinine, lipids)
- Vital sign trajectories (BMI, blood pressure)
- Missing lab values (proxy for non-engagement)
Social Determinants (177 features):
- CDC Social Vulnerability Index by zip code
- Distance to nearest oncology center
- Food desert proximity
- Public transit access to dialysis centersStage 2: Gradient Boosting Ensemble
XGBOOST + LIGHTGBMTrain XGBoost and LightGBM models on 36 months of member-month observations. Target: did member cross $100K in total paid claims within the next 12 months? Ensemble weights tuned to maximize precision at top 1% (identify highest-risk members for proactive outreach without flooding care management team with false positives).
# Model Training Configuration
Data Split:
- Training: 24 months (70% of data)
- Validation: 6 months (15% of data)
- Test: 6 months (15% of data, held-out employers)
XGBoost Hyperparameters:
max_depth: 8
learning_rate: 0.03
n_estimators: 1500
subsample: 0.8
colsample_bytree: 0.8
objective: "binary:logistic"
eval_metric: ["auc", "aucpr"]
LightGBM Hyperparameters:
num_leaves: 127
learning_rate: 0.03
n_estimators: 1500
feature_fraction: 0.8
objective: "binary"
metric: ["auc", "binary_logloss"]
Ensemble:
Final_Score = 0.55 × XGBoost + 0.45 × LightGBM
Performance on Test Set:
AUC-ROC: 0.84
AUC-PR: 0.62
Precision @ Top 1%: 68% (68 out of 100 flagged members cross $100K)
Precision @ Top 5%: 42% (actionable for care management)
False Positive Rate: 3.2% (acceptable)Stage 3: Explainability & Risk Scoring
SHAP VALUESFor each high-risk member, SHAP (SHapley Additive exPlanations) values decompose the model's prediction into feature contributions. This enables care managers to understand WHY the model flagged someone and what interventions might mitigate risk (e.g., "insulin dose escalation + 2 ER visits + missed cardiology follow-up" → diabetic crisis trajectory).
# SHAP Explainability Output
Member ID: M789456
Risk Score: 0.87 (Top 0.3% of population)
Predicted 12-Month Cost: $180K-$240K (P50-P75 range)
Top Risk Drivers:
1. +0.24 Insulin dose increased 180% in 90 days
2. +0.18 HbA1c trend: 7.2 → 9.8 → 11.3 (deteriorating control)
3. +0.15 2 ER visits with DKA diagnosis codes in 6 months
4. +0.12 Missed 3 consecutive endocrinology appointments
5. +0.09 Lives 28 miles from nearest diabetes educator
6. +0.07 Fills medications at 3 different pharmacies (fragmented care)
7. +0.05 History of noncompliance (MPR < 60% for 2+ drug classes)
Recommended Interventions:
→ Immediate outreach by diabetic care manager
→ Schedule endocrinology appointment within 7 days
→ Enroll in remote glucose monitoring program
→ Consolidate pharmacy fills (medication synchronization)
→ Address social barriers (transportation assistance to appointments)
Expected Impact of Interventions:
- Reduce 12-month cost by 30-45% ($54K-$108K savings)
- Prevent hospitalization (80% probability if no intervention)Care Management Integration
Weekly Predictive Pipeline
# Weekly Refresh Workflow
Monday 3 AM:
1. Ingest updated claims data (medical + Rx) from prior week
2. Refresh biometric data from health portal integrations
3. Re-score all active members (typically 5-20 minutes for 5K lives)
4. Rank by risk score descending
Monday 8 AM:
5. Care management team reviews Top 50 list:
- New entrants (jumped into Top 50 this week)
- Score accelerations (moved up 20+ positions)
- Deceleration (dropped out — intervention working?)
6. Automated outreach triggers:
- SMS to member: "Your care team wants to connect about your health goals"
- Email to PCP: "Member X flagged for care coordination — attached clinical summary"
- Alert to HR benefits team (if member also shows benefit non-utilization signals)
7. Care manager assigns cases:
- Tier 1 (Score > 0.80): Immediate phone outreach + home visit if needed
- Tier 2 (Score 0.60-0.80): Telephonic case management
- Tier 3 (Score 0.40-0.60): Digital nudges + educational materials
Thursday Review:
8. Track intervention outcomes:
- Did member engage with care manager?
- Was specialist appointment scheduled?
- Did clinical indicators improve (e.g., HbA1c retest)?
Monthly Retrospective:
9. Validate model performance:
- Of members predicted to cross $100K, how many actually did?
- Of interventions deployed, what was cost impact?
- Retrain model if drift detected (AUC drops below 0.80)Prioritized Outreach List
Top 50 highest-risk members updated weekly with SHAP explanations for each
Real-Time Risk Monitoring
Dashboard showing risk score trends, new high-risk entrants, and intervention status
Financial Exposure Forecast
Probabilistic total large claim spend (P50/P75/P90) for budget planning
Intervention Outcomes
Model flagged member with rapidly escalating insulin dose + missed endocrinology visits. Care manager scheduled immediate appointment, enrolled in CGM program, addressed transportation barrier. Member stabilized HbA1c from 11.8 → 7.4 over 6 months. Avoided predicted hospitalization ($45K) and long-term complications (estimated $135K over 2 years).
Model identified member with multiple imaging studies + specialist referrals across 3 different systems (fragmented care signal). Navigation nurse coordinated single multi-disciplinary visit at comprehensive cancer center. Early-stage diagnosis enabled less-aggressive treatment ($85K vs. $405K for late-stage). Member outcome significantly improved.
Model flagged expectant mother with gestational diabetes + hypertension + prior preterm delivery. Enrolled in high-risk maternity program with weekly monitoring. Delivered at 37 weeks (full-term) vs. predicted 31 weeks. Avoided NICU stay (average $12K/day × 42 days = $504K). Actual delivery cost: $54K vs. predicted $504K+.
Predict Large Claims Before They Hit Your Balance Sheet
Deploy AI-powered large claimant prediction with 6-18 month lead time. Proactive care management, Centers of Excellence steering, and dynamic stop-loss optimization—all driven by real-time ML scoring.
Request Model Demo→