Egreen Quanta is a research prototype for early disease-detection research using classical machine learning and experimental hybrid quantum machine learning techniques. The current implementation focuses on Alzheimer's disease research using the OASIS-2 longitudinal dataset.
Research prototype: The system is intended for academic experimentation and demonstration. It is not a clinically validated diagnostic system.
The platform combines OASIS-2 research data, data preparation, classical machine learning, experimental quantum machine learning, a FastAPI backend, and a React + Vite frontend.
The current live inference path uses the classical Random Forest model.
The quantum machine-learning component is maintained as an experimental
research branch and is not connected to the live /predict endpoint.
The current machine-learning pipeline uses these OASIS-2 features:
Age--- participant ageEDUC--- years of educationSES--- socioeconomic statusMMSE--- Mini-Mental State Examination scoreeTIV--- Estimated Total Intracranial VolumenWBV--- Normalized Whole Brain VolumeASF--- Atlas Scaling FactorM/F--- biological sex
The following fields are not used as model features:
Subject IDMRI IDVisitGroupCDR
The current prototype uses three research-oriented classes:
NormalMCI-likeAD
The CDR-based prototype mapping is:
CDR Prototype label
0.0 Normal
0.5 MCI-like
1.0 AD
2.0 AD
These are research label mappings used for machine-learning experimentation. They are not equivalent to clinical diagnostic criteria.
OASIS-2 Research Data
↓
Data Cleaning / Preparation
↓
Feature Filtering & Preparation
↓
Subject-Level Train/Test Split
↓
┌───────────────────────────────┐
│ Classical ML │
│ LR / SVM / Random Forest │
└───────────────────────────────┘
↓
Model Evaluation & Comparison
↓
Current Best Model: Random Forest
↓
FastAPI /predict
↓
React Frontend
↓
Prediction + Class Probabilities
The project also contains a separate experimental QML pipeline for research comparison.
Three classical baselines were evaluated:
- Logistic Regression
- Support Vector Machine (SVM)
- Random Forest
The current model-selection metric is Macro F1 because the dataset is class-imbalanced and performance across all classes is important.
The fixed evaluation split contains:
- 120 training subjects
- 30 testing subjects
- 298 training rows
- 75 testing rows
Results:
Model Accuracy Macro Precision Macro Recall Macro F1
Random Forest 74.67% 68.64% 80.37% 70.77% SVM 66.67% 60.26% 73.48% 60.95% Logistic Regression 66.67% 57.64% 64.66% 58.97%
Random Forest is the current selected model and powers the live prediction API.
The saved inference model is:
ml/best_model.joblib
The fixed holdout classification report for the current Random Forest was:
Class Precision Recall F1 Support
Normal 0.82 0.82 0.82 40 MCI-like 0.77 0.59 0.67 29 AD 0.46 1.00 0.63 6
Confusion matrix:
Predicted
Normal MCI-like AD
Actual Normal 33 5 2
MCI-like 7 17 5
AD 0 0 6
The AD test support is only 6 rows. Therefore, the 100% AD recall on this particular holdout must not be interpreted as evidence of 100% general AD detection performance.
The generated confusion-matrix artifact is:
confusion_matrix.png
Because OASIS-2 contains multiple visits for the same subjects, cross-validation was also performed at the subject level.
Five-fold cross-validation produced the following mean results:
Model Mean Accuracy Mean Macro F1 Accuracy Std Macro F1 Std
Random Forest 0.6653 0.5870 0.0354 0.0827 SVM 0.5931 0.5161 0.0472 0.0286 Logistic Regression 0.6768 0.5756 0.0560 0.0793
Mean per-class F1 for the original Random Forest:
Class Mean F1
Normal 0.7787 MCI-like 0.5337 AD 0.4485
The Random Forest remains the preferred model when Macro F1 is used as the selection criterion, although Logistic Regression has slightly higher mean accuracy in cross-validation.
Random Forest impurity-based feature importance was also examined.
Feature Importance
MMSE 27.51% nWBV 17.24% Age 13.59% eTIV 13.14% ASF 12.69% EDUC 7.38% SES 4.23% M/F ~4.22% combined
MMSE is the largest individual feature by this importance measure.
The MRI-derived quantitative features nWBV, eTIV, and ASF together
account for about 43.07% of the Random Forest impurity importance. This
is a relative model-importance measure and should not be interpreted as
a percentage of disease causation, diagnostic contribution, or clinical
risk.
Artifacts:
ml/feature_importance.csv
ml/feature_importance.png
Random Forest hyperparameter tuning was investigated using subject-level cross-validation.
The investigated tuned configuration was:
n_estimators = 300
max_depth = None
max_features = sqrt
min_samples_leaf = 2
The tuned configuration achieved approximately:
Mean cross-validation Macro F1 = 0.6160
However, it did not improve the fixed holdout:
Model Holdout Accuracy Holdout Macro F1
Original Random Forest 74.67% 70.77% Tuned Random Forest 73.33% 69.55%
Therefore, the original Random Forest was retained as the live inference model.
Research conclusion:
Hyperparameter tuning was investigated using subject-level cross-validation. Although the tuned configuration improved mean cross-validation Macro-F1, it did not improve performance on the fixed holdout set; therefore, the original Random Forest was retained as the current inference model.
OASIS-2 is longitudinal, so an additional experimental experiment investigated whether features from an earlier visit could predict whether CDR would worsen at the subsequent visit.
The experiment:
- constructed consecutive visit pairs
- used features from the earlier visit only
- defined the target as future CDR > current CDR
- kept all visits from a subject in the same train/test partition
- used a balanced Random Forest baseline
Results on the fixed test split:
Metric Result
Accuracy 76.92% Precision for worsening 20.00% Recall for worsening 16.67% F1 for worsening 18.18%
There were only 6 worsening cases in the test portion.
Therefore, this experiment is not used for deployment and should not be presented as a reliable early-progression predictor.
The experiment is implemented in:
ml/progression_baseline.py
The project contains an experimental quantum machine-learning pipeline implemented with PennyLane.
8 input features
↓
M/F → numeric encoding
↓
Median imputation
↓
StandardScaler
↓
PCA → 4 components
↓
4 qubits
↓
RY data encoding
↓
3 trainable Rot layers
↓
Neighboring CNOT entanglement
↓
Pauli-Z expectation values
↓
Classical output layer
↓
Softmax class probabilities
The QML experiments use the PennyLane default.qubit simulator.
The PCA step retains approximately 86.17% of the training variance in the four components.
The current 40-epoch hybrid QML experiment achieved approximately:
- Accuracy: 50.67%
- Macro F1: approximately 34%
A 100-epoch experiment was also performed:
- Accuracy: 46.67%
- Macro F1: approximately 37%
Although training loss continued to decrease with more epochs, test performance did not improve materially.
A simpler quantum feature-extractor experiment was also evaluated:
- Accuracy: 37.33%
- Macro F1: 33%
A classical Logistic Regression model using the same PCA(4) representation achieved:
- Accuracy: 66.67%
- Macro F1: 58.97%
This comparison indicates that the current QML circuit/training setup is the main experimental bottleneck rather than PCA compression alone.
No quantum advantage is claimed. The QML component is research-only and is not connected to the live prediction API.
The frontend includes a research-oriented explainability view for the live Random Forest prediction.
The current explanation method is a local baseline-replacement sensitivity analysis:
- Obtain the original predicted-class probability.
- Replace one feature at a time with a fitted preprocessing baseline.
- Recalculate the predicted-class probability.
- Use the probability change as the local feature contribution.
Interpretation:
- Positive contribution: replacing the feature with its baseline decreases the predicted-class probability, so the original feature supported the prediction.
- Negative contribution: replacing the feature with its baseline increases the predicted-class probability, so the original feature opposed the prediction.
These values are not causal effects, calibrated confidence values, or additive feature attributions. The implementation is not SHAP.
Advanced methods such as SHAP-based interpretation remain future work.
The backend uses FastAPI.
Health check:
GET /health
Prediction:
POST /predict
The prediction endpoint accepts:
Age
EDUC
SES
MMSE
eTIV
nWBV
ASF
M_F
and returns:
- predicted prototype class
- class probabilities
- local feature contributions
The live inference path is:
React → FastAPI /predict → ml/predict.py → Random Forest → response
The QML experiments are not part of this live path.
The frontend uses React + Vite.
Current research-oriented pages include:
- Clinical Hub
- Benchmarks
- Quantum Circuit
- Explainability
The interface presents:
- patient feature inputs
- live Random Forest inference
- class probabilities
- prototype class visualization
- model benchmark results
- experimental QML architecture
- local feature-contribution research output
- research/prototype status and limitations
OASIS-2
↓
Data Preparation
↓
Feature Filtering & Preparation
↓
Subject-Level Split
↓
┌────────────┴────────────┐
↓ ↓
Classical ML Experimental QML
LR / SVM / RF StandardScaler
↓ PCA → 4 components
Model Evaluation 4-qubit VQC
↓ ↓
Current Best: RF Research Comparison
↓
FastAPI
↓
React UI
↓
Prediction / Probabilities / Explainability
egreen-quanta/
├── backend/
│ └── app/
│ └── main.py
├── frontend/
│ └── src/
│ ├── components/
│ ├── layout/
│ └── pages/
├── ml/
│ ├── train.py
│ ├── predict.py
│ ├── evaluate.py
│ ├── cv_evaluate.py
│ ├── feature_importance.py
│ ├── compare_tuned_rf.py
│ ├── progression_baseline.py
│ ├── tune_rf.py
│ └── qml/
│ ├── quantum_model.py
│ ├── hybrid_model.py
│ └── qml_training_test.py
├── qml_test.py
├── .gitignore
└── README.md
Generated evaluation artifacts such as confusion_matrix.png and
feature-importance files may be kept locally depending on the intended
repository presentation.
Place the OASIS-2 research dataset in the local raw-data directory. The raw dataset is intentionally excluded from version control.
Example:
data/raw/oasis_longitudinal_demographics.xlsx
From the repository root:
python3 ml/train.py \
--data data/raw/oasis_longitudinal_demographics.xlsx \
--output mlThis evaluates the classical baselines and saves the selected model as:
ml/best_model.joblib
python3 ml/evaluate.py \
--data data/raw/oasis_longitudinal_demographics.xlsx \
--model ml/best_model.joblibpython3 ml/qml/quantum_model.pyThe hybrid QML implementation is in:
ml/qml/hybrid_model.py
From the repository root:
python3 -m uvicorn backend.app.main:app --reloadThe API will be available at:
http://127.0.0.1:8000
FastAPI documentation:
http://127.0.0.1:8000/docs
cd frontend
npm install
npm run devVite will provide the local frontend URL in the terminal, normally:
http://localhost:5173
The project currently uses the OASIS-2 longitudinal dataset for research experimentation.
The dataset contains multiple visits for subjects, which is why subject-level partitioning is used to reduce the risk of placing different visits from the same subject into both training and testing sets.
The raw research dataset is intentionally excluded from version control
through .gitignore.
- OASIS-2 data preparation
- CDR-based prototype label mapping
- subject-level train/test splitting
- classical ML baseline comparison
- Random Forest model selection
- fixed-holdout evaluation
- confusion-matrix evaluation
- subject-level cross-validation
- Random Forest feature-importance analysis
- Random Forest hyperparameter-tuning experiment
- experimental progression-risk baseline
- experimental QML pipeline
- QML epoch comparison
- same-PCA classical comparison
- FastAPI prediction endpoint
- React frontend integration
- live class probabilities
- research-oriented explainability
- research benchmark UI
- quantum-circuit research UI
Random Forest is the current live inference model.
- Hybrid QML
- progression-risk prediction
- local feature-contribution analysis
These components are not presented as clinically validated systems.
- Python
- pandas
- scikit-learn
- PennyLane
- FastAPI
- Pydantic
- React
- Vite
- JavaScript
- Git
- GitHub
OASIS-2 contains repeated visits from the same subjects. Train/test and cross-validation partitions are therefore made at the subject level to reduce information leakage.
The classical pipeline uses:
- median imputation for numeric features
- StandardScaler for numeric features
- most-frequent imputation for
M/F - one-hot encoding for
M/F
The project uses feature filtering & preparation, not a formal automated feature-selection algorithm.
Macro F1 is used as the primary model-selection metric so that performance across the three prototype classes is considered rather than relying only on overall accuracy.
This project is currently a research and demonstration prototype.
Important limitations include:
- The model has not been clinically validated.
- OASIS-2 is a research dataset and does not represent all populations.
- The prototype labels are based on CDR mappings and are not equivalent to clinical diagnostic criteria.
- The current feature-based pipeline does not perform direct end-to-end MRI image classification.
- Model probabilities should not be interpreted as clinical confidence.
- The fixed holdout contains only 6 AD rows.
- The progression experiment has very few positive worsening cases and performed poorly on worsening recall/F1.
- The QML component is experimental and currently underperforms the classical Random Forest baseline.
- No quantum advantage has been demonstrated.
- Larger, more diverse datasets and independent validation are required.
- The current explainability method is local sensitivity analysis rather than a formal causal or SHAP attribution method.
Potential future improvements include:
- MRI image-based deep learning
- advanced MRI feature extraction
- SHAP-based model interpretation
- stronger explainability methods
- additional validation strategies
- larger and more diverse datasets
- external validation
- improved hybrid quantum-classical architectures
- QML ablation studies
- quantum hardware experimentation
- progression modeling with substantially more longitudinal data
- calibration and uncertainty analysis
- model comparison across additional algorithms
- improved clinical research evaluation
Egreen Quanta is an academic/research prototype developed for experimentation with machine learning and hybrid quantum machine learning.
It is not intended for medical diagnosis, clinical decision-making, or treatment recommendations.
Model predictions and probabilities should not be considered medical advice or clinical confidence.