Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for HealthAI

Model Card: Alzheimer's Disease MRI Classification System

Model Name: AlzheimerNet-ResNet18 Version: 1.0 Last Updated: December 20, 2025 Model Type: Convolutional Neural Network (Transfer Learning) Task: Multi-class Image Classification (Medical Imaging)

---

Model Details

Overview

A deep learning model for automated classification of Alzheimer's disease stages from brain MRI scans. Based on ResNet18 architecture with transfer learning from ImageNet, adapted for grayscale medical imaging.

Intended Use

• Primary Use: Automated screening and classification support for Alzheimer's disease diagnosis
• Target Users: Healthcare professionals, radiologists, neurologists, clinical researchers
• Use Context: Clinical decision support tool for analyzing structural brain MRI scans
• NOT INTENDED FOR: Standalone diagnosis without physician oversight, replacement of clinical judgment, or use by non-medical professionals

Model Architecture

• Base Architecture: ResNet18 (He et al., 2016)
• Pre-training: ImageNet (natural images)
• Modifications:
  • Input adapted from RGB (3 channels) to grayscale (1 channel)
  • Output layer modified for 4-class classification
  • All layers fine-tuned on medical data
• Parameters: ~11 million trainable parameters
• Input: 224×224 grayscale MRI images, normalized to [-1, 1]
• Output: Probability distribution over 4 classes (softmax)

Training Details

• Training Data: 5,120 labeled brain MRI scans
• Validation Data: 1,280 samples (20% holdout)
• Test Data: 1,280 unlabeled samples
• Data Augmentation: Rotation (±10°), horizontal flip, brightness/contrast jitter
• Optimizer: Adam (learning rate: 1e-4, weight decay: 1e-5)
• Training Duration: 20 epochs (~90 minutes on Tesla T4 GPU)
• Batch Size: 32
• Loss Function: Cross-entropy

Performance Metrics

Metric	Value
Validation Accuracy	97.27%
Training Accuracy	99.19%
Precision (weighted avg)	97%
Recall (weighted avg)	97%
F1-Score (weighted avg)	97%

Per-Class Performance:

• Class 0 (Non-Demented): 98% F1-score
• Class 1 (Very Mild): 100% F1-score
• Class 2 (Mild): 98% F1-score
• Class 3 (Moderate): 96% F1-score

---

Limitations

Data Limitations

1. Limited Sample Size

   • Training set of 5,120 samples is modest for deep learning
   • May not capture full variability of real-world clinical presentations
   • Impact: Reduced generalization to rare or unusual cases

2. Class Imbalance

   • Class 1 (Very Mild) severely underrepresented (only 10 validation samples)
   • Class 2 (Mild) dominates dataset (~50% of samples)
   • Impact: Model may be less reliable for detecting very mild cases; perfect validation performance (100%) on Class 1 may not generalize

3. Single Data Source

   • All training data appears to come from similar scanner/protocol
   • No diversity in acquisition parameters, scanner manufacturers, or imaging protocols
   • Impact: May not generalize well to different MRI scanners or imaging centers

4. Temporal Snapshot

   • Uses single time-point imaging only
   • No longitudinal progression data
   • Impact: Cannot predict disease trajectory or rate of decline

Model Limitations

1. Black Box Nature

   • Deep learning model lacks full interpretability
   • Difficult to explain specific predictions to clinicians
   • Impact: May reduce trust and clinical adoption

2. Overfitting Risk

   • 99.19% training accuracy suggests potential memorization
   • Small train-val gap (2%) is positive, but vigilance needed
   • Impact: Performance may degrade on truly novel cases

3. Binary Thresholding

   • Provides discrete class predictions rather than continuous severity scores
   • Real disease progression is continuous
   • Impact: May miss subtle transitions between stages

4. No Uncertainty Quantification

   • Model doesn't provide confidence intervals or prediction uncertainty
   • All predictions treated equally regardless of confidence
   • Impact: Cannot flag ambiguous cases for manual review

Technical Constraints

1. Computational Requirements

   • Requires GPU for inference (CPU too slow for clinical deployment)
   • Model size (~45MB) manageable but not edge-deployable
   • Impact: Limits deployment to well-resourced facilities

2. Input Format Constraints

   • Requires specific image preprocessing (224×224 resize, normalization)
   • Sensitive to image quality and artifacts
   • Impact: May fail on low-quality or corrupted scans

3. Single Modality

   • Uses structural MRI only
   • Ignores functional MRI, PET, CSF biomarkers, genetics, cognitive scores
   • Impact: Misses complementary diagnostic information

---

Bias & Fairness Considerations

Potential Sources of Bias

1. Demographic Bias (Unknown)

   • Issue: Dataset demographics (age, sex, race, ethnicity, socioeconomic status) not documented
   • Risk: Model may perform differently across demographic groups
   • Example: If training data over-represents Caucasian populations, may underperform on other ethnicities
   • Mitigation Needed: Demographic analysis of training data and subgroup performance evaluation

2. Selection Bias

   • Issue: Dataset may not represent general population (e.g., clinical trial participants vs. real-world patients)
   • Risk: Higher prevalence of severe cases or younger patients in research datasets
   • Impact: May misclassify community-dwelling, less severe cases
   • Mitigation: Validate on diverse, real-world clinical populations

3. Scanner & Protocol Bias

   • Issue: Training data likely from limited scanner types/imaging protocols
   • Risk: Performance degradation on scans from different equipment or settings
   • Impact: Model may favor specific MRI characteristics over disease features
   • Mitigation: Multi-site validation with heterogeneous scanners

4. Labeling Bias

   • Issue: Ground truth labels based on clinical diagnosis, which has inherent subjectivity
   • Risk: Model learns clinician biases rather than objective disease features
   • Impact: May perpetuate diagnostic disparities
   • Mitigation: Multiple expert consensus labels, neuropathological confirmation

5. Socioeconomic Bias

   • Issue: Access to MRI scans correlates with socioeconomic status
   • Risk: Underrepresentation of lower-income populations in training data
   • Impact: May not generalize to underserved communities
   • Mitigation: Diverse data collection from community health centers

Fairness Metrics

Current Status: ⚠️ Not Evaluated

Required Analysis:

• [ ] Stratified performance by age groups (60-70, 70-80, 80+)
• [ ] Stratified performance by biological sex (if known)
• [ ] Stratified performance by race/ethnicity (if known)
• [ ] Error rate disparity across subgroups
• [ ] False positive/negative rate parity
• [ ] Equal opportunity metrics

Recommendation: Before clinical deployment, conduct comprehensive fairness audit with demographic-stratified evaluation.

Ethical Considerations

1. False Positives (Type I Error)

   • Impact: Unnecessary patient anxiety, costly follow-up testing
   • Current Rate: ~3% overall (varies by class)
   • Clinical Consequence: Mild - requires confirmatory testing anyway

2. False Negatives (Type II Error)

   • Impact: Missed early diagnosis, delayed treatment
   • Current Rate: ~3% overall (varies by class)
   • Clinical Consequence: SEVERE - early intervention critical for AD
   • Mitigation: Tune threshold to favor sensitivity over specificity if used for screening

3. Automation Bias

   • Risk: Clinicians may over-rely on model predictions
   • Impact: Reduced clinical judgment, missed complex cases
   • Mitigation: Emphasize model as decision support, not replacement

4. Data Privacy

   • Risk: MRI scans are protected health information (PHI)
   • Impact: HIPAA violations, patient privacy breaches
   • Mitigation: De-identification, secure storage, limited access

---

Interpretability

Current Interpretability: ⚠️ Limited (Black Box)

What We Can Interpret:

1. Class Predictions

   • Model outputs clear class labels (0-3)
   • Softmax probabilities indicate relative confidence
   • Limitation: Doesn't explain why

2. Confusion Patterns

   • Most errors between Class 2 ↔ Class 3 (adjacent stages)
   • Clinically plausible confusion (subtle differences)
   • Insight: Model learns clinically relevant feature boundaries

3. Feature Learning (Abstract)

   • Early layers detect edges, textures (brain structure)
   • Middle layers detect anatomical patterns (ventricles, cortex)
   • Late layers detect disease signatures (atrophy, enlargement)
   • Limitation: Specific features not directly visible

What We CANNOT Interpret:

1. Spatial Attribution

   • Which brain regions drive each prediction?
   • Are decisions based on hippocampus, cortex, ventricles, or multiple areas?
   • Missing: Saliency maps, attention weights, GradCAM visualizations

2. Decision Boundaries

   • What specific features distinguish Class 2 from Class 3?
   • How much atrophy is "enough" for severe classification?
   • Missing: Feature importance scores, counterfactual examples

3. Individual Predictions

   • Why was this specific patient classified as Class 3?
   • Missing: Case-by-case explanations

Recommended Interpretability Enhancements:

High Priority:

1. GradCAM/GradCAM++ - Highlight influential brain regions
2. Attention Mechanisms - Built-in interpretability through attention weights
3. Saliency Maps - Pixel-level importance visualization

Medium Priority:

1. Feature Visualization - Show what specific neurons detect
2. Layer-wise Relevance Propagation (LRP) - Trace predictions back to inputs
3. SHAP Values - Local feature importance

Low Priority (Research):

1. Concept Activation Vectors - High-level semantic concepts
2. Prototypical Examples - Show similar training cases

Clinical Interpretability Requirements:

For clinical adoption, we need to provide:

• ✅ Prediction confidence scores (currently available via softmax)
• ❌ Brain region heatmaps (NOT IMPLEMENTED)
• ❌ Comparison to "typical" cases (NOT IMPLEMENTED)
• ❌ Uncertainty quantification (NOT IMPLEMENTED)
• ❌ Explanation of decision (NOT IMPLEMENTED)

Status: Model currently unsuitable for clinical deployment without interpretability enhancements.

---

Out-of-Scope Use Cases

Explicitly NOT INTENDED FOR:

1. ❌ Standalone Clinical Diagnosis

   • Model must be used as decision support ONLY
   • Requires confirmation by qualified healthcare professionals
   • Not a replacement for comprehensive clinical evaluation

2. ❌ Predictive Prognosis

   • Cannot predict future disease progression or survival
   • Not trained on longitudinal outcome data

3. ❌ Treatment Recommendation

   • Does not suggest specific treatments or interventions
   • Clinical management decisions require physician expertise

4. ❌ Non-MRI Modalities

   • Trained exclusively on structural MRI
   • Will fail on CT, PET, ultrasound, or X-ray images

5. ❌ Pediatric or Non-AD Dementia

   • Trained on adult Alzheimer's disease only
   • Not applicable to frontotemporal dementia, Lewy body dementia, vascular dementia, etc.

6. ❌ Real-Time Critical Decisions

   • Not validated for emergency or time-sensitive scenarios
   • Requires proper quality control and validation

7. ❌ Consumer/Direct-to-Patient Use

   • Requires medical expertise to interpret
   • Not designed for self-diagnosis

---

Caveats & Recommendations

Deployment Considerations

1. Regulatory Approval Required

   • Not FDA-cleared or CE-marked
   • Requires validation for medical device classification
   • Must comply with local healthcare regulations

2. Clinical Validation Needed

   • External validation on independent datasets
   • Prospective clinical trial to assess real-world performance
   • Comparison to radiologist performance

3. Quality Control

   • Implement input validation (image quality checks)
   • Monitor prediction drift over time
   • Regular re-validation as new data emerges

4. Human Oversight Mandatory

   • All predictions require physician review
   • System should flag uncertain predictions
   • Maintain audit trail of predictions vs. final diagnoses

Safe Use Guidelines

DO:

• ✅ Use as screening tool to prioritize cases
• ✅ Validate predictions with clinical assessment
• ✅ Monitor performance on your local population
• ✅ Retrain periodically with new data
• ✅ Document all model decisions

DON'T:

• ❌ Use without physician oversight
• ❌ Apply to populations not represented in training data
• ❌ Ignore model uncertainty or low confidence predictions
• ❌ Deploy without local validation
• ❌ Use for legal or financial decisions

---

Model Versioning & Updates

Current Version: 1.0 (Baseline)

• Release Date: December 20, 2025
• Training Data Version: Kaggle MRI Alzheimer's Dataset (Dec 2025)
• Performance: 97.27% validation accuracy

Planned Updates:

Version 1.1 (Proposed - Q1 2026)

• Implement GradCAM interpretability
• Add uncertainty quantification
• Address Class 1 imbalance with synthetic augmentation

Version 2.0 (Proposed - Q2 2026)

• Multi-site validation
• Ensemble model for improved robustness
• Demographic fairness audit and mitigation

---

Contact & Feedback

Model Developers: [Your Name/Team] Institution/Organization: [Your Organization] Email: [Contact Email] Issues & Feedback: [GitHub Issues / Email]

Reporting Errors or Concerns

If you encounter:

• Unexpected predictions or errors
• Bias or fairness issues
• Safety concerns
• Technical bugs

Please contact us immediately with:

• Anonymized case details
• Input image characteristics
• Expected vs. actual output
• Your use context

---

Acknowledgments

• AI for Alzheimer's Hackathon organizers
• Dataset providers and contributors
• Open-source PyTorch and torchvision communities
• Medical imaging research community

---

License & Terms of Use

License: [To Be Determined - specify open-source or proprietary]

Terms:

• Research and educational use permitted
• Clinical use requires additional validation and regulatory approval
• Commercial use requires separate licensing agreement
• No warranties provided - use at your own risk
• Users assume all liability for clinical decisions

---

Changelog

Version 1.0 (December 20, 2025)

• Initial release
• ResNet18 baseline model
• 97.27% validation accuracy
• 4-class Alzheimer's classification
• Known limitations documented

---

This model card follows guidelines from Mitchell et al. (2019) "Model Cards for Model Reporting" and the EU AI Act technical documentation requirements.

Last Updated: December 20, 2025 Next Review: March 20, 2026 (quarterly review)

Built With

• numpy
• pandas
• pil
• scikit-learn
• seaborn
• torch
• torchvision