# Model Cards

Detailed documentation for each Presage model, following the Model Card framework for transparent AI documentation.

---

## presage-vitals-v1

### Model Details

| | |
|---|---|
| **Name** | Presage Vitals Risk Assessment |
| **Version** | 1.2.0 |
| **Type** | Gradient Boosted Trees + Neural Network Ensemble |
| **Released** | March 2024 |
| **Last Updated** | November 2024 |

### Intended Use

**Primary Use**: Predict patient deterioration risk within 24 hours based on vital signs.

**Intended Users**:
- Healthcare providers in hospital settings
- Clinical decision support systems
- Remote patient monitoring platforms

**Out of Scope**:
- Diagnostic decisions without clinician oversight
- Pediatric patients (< 18 years) - separate model available
- Outpatient/home settings without clinical supervision

### Input Requirements

```json
{
  "heart_rate": 72,              // Required, beats/min (30-250)
  "blood_pressure_systolic": 120, // Required, mmHg (60-260)
  "blood_pressure_diastolic": 80, // Required, mmHg (30-160)
  "respiratory_rate": 16,         // Required, breaths/min (4-60)
  "temperature": 98.6,            // Required, °F (90-110)
  "oxygen_saturation": 98,        // Required, % (50-100)
  "age": 45,                      // Optional, years
  "consciousness_level": "alert"  // Optional: alert/verbal/pain/unresponsive
}
```

### Output

```json
{
  "risk_score": 0.12,      // 0.0-1.0 continuous score
  "risk_level": "low",     // low/moderate/high/critical
  "confidence": 0.94,      // Model confidence
  "contributing_factors": [
    {"factor": "respiratory_rate", "impact": "positive", "magnitude": 0.3}
  ]
}
```

### Performance Metrics

Evaluated on held-out test set (n=25,000 patients):

| Metric | Value | 95% CI |
|--------|-------|--------|
| AUROC | 0.89 | 0.87-0.91 |
| AUPRC | 0.45 | 0.42-0.48 |
| Sensitivity @ 90% Specificity | 0.85 | 0.82-0.88 |
| Calibration Error (ECE) | 0.03 | - |

### Limitations

- **Trained on US hospital data**: May have reduced performance on populations with different baseline health profiles
- **Vital signs only**: Does not incorporate lab values, medications, or medical history
- **Snapshot predictions**: Based on single timepoint; trends not considered
- **Not validated for**: ICU patients, post-surgical patients, pregnant patients

### Fairness Evaluation

| Subgroup | AUROC | Sensitivity |
|----------|-------|-------------|
| Male | 0.89 | 0.85 |
| Female | 0.88 | 0.84 |
| Age 18-40 | 0.87 | 0.82 |
| Age 41-65 | 0.89 | 0.86 |
| Age 65+ | 0.90 | 0.87 |

### Training Data

- **Source**: De-identified EHR data from 12 US hospitals
- **Size**: 2.1 million patient encounters
- **Period**: 2018-2023
- **Labels**: 24-hour deterioration events (ICU transfer, rapid response, death)

---

## presage-clinical-v1

### Model Details

| | |
|---|---|
| **Name** | Presage Clinical Decision Support |
| **Version** | 2.0.1 |
| **Type** | Transformer-based Neural Network |
| **Released** | June 2024 |
| **Last Updated** | January 2025 |

### Intended Use

**Primary Use**: Provide clinical decision support by analyzing patient data and suggesting relevant considerations.

**Intended Users**:
- Emergency department physicians
- Hospitalists
- Advanced practice providers

**Out of Scope**:
- Autonomous decision-making
- Replacing clinical judgment
- Legal or billing documentation

### Input Requirements

```json
{
  "chief_complaint": "chest pain",
  "vitals": { ... },
  "demographics": {
    "age": 55,
    "sex": "male"
  },
  "medical_history": ["hypertension", "diabetes"],
  "medications": ["metformin", "lisinopril"],
  "recent_labs": { ... }  // Optional
}
```

### Output

```json
{
  "considerations": [
    {
      "condition": "Acute Coronary Syndrome",
      "likelihood": "moderate",
      "suggested_workup": ["ECG", "Troponin", "Chest X-ray"],
      "reasoning": "55yo male with HTN, DM presenting with chest pain"
    }
  ],
  "risk_flags": ["cardiovascular_risk_factors"],
  "confidence": 0.87
}
```

### Performance Metrics

| Metric | Value |
|--------|-------|
| Agreement with Expert Panel | 91% |
| Critical Diagnosis Sensitivity | 97% |
| Specificity for Low-Risk Discharge | 89% |

### Limitations

- **English language only**
- **US clinical practice patterns**: Recommendations reflect US standard of care
- **Not real-time**: Does not monitor for changes after initial assessment
- **Requires structured input**: Free-text notes must be preprocessed

---

## presage-diagnostic-v1

### Model Details

| | |
|---|---|
| **Name** | Presage Diagnostic Assistant |
| **Version** | 0.9.0 (Beta) |
| **Type** | Multi-modal Neural Network |
| **Released** | Beta - Limited Access |

### Status

This model is currently in **beta** and available only to select research partners. Clinical trial in progress (NCT04ZZZZZZ).

### Intended Use

Multi-condition diagnostic support combining clinical data, lab values, and imaging reports.

### Access

Contact research@presage.tech for beta access criteria.

---

## Requesting Updates

For questions about model cards or to report issues:

- Technical: support@presage.tech
- Clinical: clinical@presage.tech
- Research: research@presage.tech