# Model Card: Dental Implant Risk Prediction

## Model Details

- **Model name**: Dental Implant Unfavorable Outcome Predictor
- **Model version**: 2.0
- **Model type**: XGBoost (gradient-boosted trees)
- **Baseline model**: L2-regularized Logistic Regression
- **Framework**: scikit-learn 1.x, XGBoost 2.x
- **Task**: Binary classification (unfavorable outcome: implant loss or peri-implantitis)
- **Input**: 25 features (14 clinical predictors after one-hot encoding)
- **Output**: Probability of unfavorable outcome [0, 1]
- **Developer**: Bakhrom Isomadinov, CAU School of Dentistry
- **Date**: 2026
- **License**: Academic use only

## Intended Use

- **Primary use**: Clinical decision support for dental implant risk assessment
- **Intended users**: Dental clinicians, implantologists, oral surgeons
- **Out-of-scope uses**:
  - NOT intended as sole basis for clinical decisions
  - NOT validated for pediatric populations (age < 18)
  - NOT validated for populations outside Taiwanese demographics
  - NOT intended for regulatory or insurance purposes

## Training Data

- **Source**: Zenodo open-access dataset (doi:10.5281/zenodo.1227714)
- **Size**: 747 implant-level records
- **Events**: 117 unfavorable outcomes (15.66%)
- **Split**: Train 522 (70%) / Validation 112 (15%) / Test 113 (15%)
- **Stratification**: By outcome to preserve class balance
- **Geographic origin**: Taiwan

## Evaluation Results

### Discrimination & Calibration
- **Logistic Regression**: AUROC=0.6477 (95% CI: 0.4955-0.7996), PR-AUC=0.2445, Brier=0.2396, Cal. Slope=0.81
- **XGBoost**: AUROC=0.7798 (95% CI: 0.6629-0.8825), PR-AUC=0.5144, Brier=0.1565, Cal. Slope=0.97

### Advanced Analysis
- AUROC Difference (XGB - LR): 0.1307 (-0.0239-0.2776)
- P-value (one-sided): 0.047
- XGB Youden Threshold: 0.3649 (J=0.4544)
- LR Youden Threshold: 0.526 (J=0.4304)
- XGB 10-fold CV AUROC: 0.7572 (+/- 0.0763)

### Fairness Across Subgroups

| Subgroup | N | Events | Prevalence | AUROC | Brier |
|----------|---|--------|------------|-------|-------|
| Overall | 113 | 18 | 0.1593 | 0.7798 | 0.1565 |
| Sex: female | 53 | 9 | 0.1698 | 0.7172 | 0.1948 |
| Sex: male | 59 | 9 | 0.1525 | 0.83 | 0.1241 |
| Age <=40 | 20 | 1 | 0.05 | 0.9474 | 0.1502 |
| Age 41-55 | 45 | 6 | 0.1333 | 0.8077 | 0.1491 |
| Age 56-70 | 38 | 8 | 0.2105 | 0.6792 | 0.1835 |
| Age >70 | 9 | 3 | 0.3333 | nan | nan |
| Jaw: mandible | 61 | 7 | 0.1148 | 0.9418 | 0.1219 |
| Jaw: maxilla | 52 | 11 | 0.2115 | 0.6419 | 0.1971 |
| Smoking: current | 9 | 2 | 0.2222 | nan | nan |
| Smoking: never | 97 | 15 | 0.1546 | 0.7833 | 0.1574 |
| Smoking: unknown | 5 | 1 | 0.2 | nan | nan |

## Features

### Input Predictors
| Category | Features |
|----------|----------|
| Demographics | age, sex |
| Behavioral | smoking_status, alcohol, betel_nut |
| Medical | diabetes_status, bisphosphonate_use, immunosuppressant_use |
| Dental history | periodontal_history |
| Anatomical | jaw, region |
| Surgical | augmentation, sinus_lift |
| Implant | implant_length_mm, implant_diameter_mm |
| Clinical | bone_density_code, surgeon_experience |

### Feature Importance (SHAP)
Top 5 features by mean |SHAP| value determine model predictions. See
`outputs/figures/shap_summary_xgboost.png` for the full beeswarm plot.

## Limitations

1. **Small sample size** (n=747): Wide confidence intervals, limited generalizability
2. **Single-center data**: Taiwanese population may not generalize globally
3. **Retrospective design**: Subject to selection and information bias
4. **No temporal validation**: Model not tested on future/held-out time periods
5. **No external validation**: Performance on independent datasets unknown
6. **Class imbalance**: 15.7% event rate with EPV ~4.7 (below recommended 10)
7. **Missing data**: Some predictors have >5% missingness, handled by imputation
8. **No follow-up data**: `followup_months` entirely missing in Zenodo dataset
9. **periodontal_history**: Not available in Zenodo, coded as empty string

## Ethical Considerations

- **Transparency**: SHAP explanations provided for every prediction
- **Fairness**: Subgroup analysis across sex and age groups conducted
- **Privacy**: All patient IDs de-identified (SHA-256 hashing)
- **Non-causal language**: All SHAP narratives use associational language
- **Clinical override**: Model predictions are advisory, not prescriptive
- **Data source**: Open-access dataset with documented provenance

## Quantitative Analyses

- Bootstrap confidence intervals (1000 iterations)
- Net Reclassification Improvement (NRI)
- Decision Curve Analysis (clinical utility)
- 10-fold cross-validation stability
- Permutation importance (model-agnostic)
- Post-hoc Platt calibration

## Caveats and Recommendations

1. **Always combine with clinical judgment** - this model is an aid, not a replacement
2. **Validate externally** before any clinical deployment
3. **Monitor for drift** if deployed in a different population
4. **Update regularly** as new data becomes available
5. **Do not use for populations** outside the training demographics without validation

---

*Generated automatically by the dental implant prediction pipeline.*
*Following Mitchell et al. (2019) Model Cards for Model Reporting.*