NeuroNova

Live demo: neuronova-alpha.vercel.app

A web-based ADHD screening tool that predicts likelihood of ADHD referral using only information available before any clinical assessment. Built for school counsellors, GPs, and parents, not clinicians. The output is always a referral recommendation, never a diagnosis.

The core motivation is equity: girls and younger children are chronically underdiagnosed with ADHD. By surfacing risk earlier, referrals can happen faster and more fairly.

---

Results

Metric	Score
F1	0.881
ROC-AUC	0.951
Accuracy	0.91
Precision (ADHD)	0.90
Recall (ADHD)	0.86

Fairness audit

Group	N	ADHD %	F1	ROC-AUC
Male	72	22.2%	0.867	0.916
Female	118	48.3%	0.885	0.947
Age < 10	66	48.5%	0.936	0.984
Age 10-13	75	36.0%	0.909	0.965
Age 14+	49	28.6%	0.692	0.888

No meaningful sex bias detected. Adolescent (14+) performance is weaker, a known limitation documented in the UI and API response.

---

Model

Soft voting ensemble of three individually calibrated pipelines:

• CalibratedClassifierCV(RandomForestClassifier, method='sigmoid', cv=5)
• CalibratedClassifierCV(ExtraTreesClassifier, method='sigmoid', cv=5)
• CalibratedClassifierCV(SVC(kernel='rbf'), method='sigmoid', cv=5)

Each pipeline includes a ColumnTransformer preprocessor (median imputation + standard scaling for numerics, one-hot encoding for handedness). Calibration is applied per sub-estimator before ensembling. Wrapping the VotingClassifier itself would strip the preprocessors and break inference on string inputs.

GridSearchCV used PredefinedSplit so the validation fold is identical across all models, making F1 scores directly comparable. class_weight='balanced' on all estimators.

XGBoost was evaluated and dropped as it overfits aggressively on 700 samples and 6 features. SHAP was replaced with permutation importance, which is compatible with VotingClassifier.

---

Dataset

ADHD-200 phenotypic dataset - 973 participants, multi-site research data. After dropping 26 pending diagnoses: 947 rows, 38.2% ADHD / 61.8% control.

Features used — only information available before clinical assessment:

Feature	Detail
Age	Child's age at screening
Sex	Male / Female
Handedness	Left / Right / Mixed (cleaned from EHI floats + site-specific codes)
Inattentive score	Conners T-score (9-90), mapped from raw form sum (0-12)
Hyper/Impulsive score	Conners T-score (9-90), mapped from raw form sum (0-12)

IQ, medication status, formal clinical scores, and secondary diagnoses were all excluded. These require clinical input and would not be available at the point of referral.

The form collects raw item sums (4 questions x 0-3). The API maps these to T-scores before passing to the model: T = 9 + (raw / 12) * 81.

Permutation importance:

Feature	Importance
Inattentive	0.311 +/- 0.037
Hyper/Impulsive	0.175 +/- 0.025
Age	0.017 +/- 0.009
Sex	0.013 +/- 0.012
Handedness	0.002 +/- 0.011

---

Stack

Backend: FastAPI · scikit-learn · joblib · uvicorn · Pydantic

Frontend: React · Vite · recharts · CSS Modules · react-router-dom · axios

Tooling: uv · ruff · pytest

---

Project structure

neuronova/
├── app/
│   ├── main.py          # FastAPI app, /predict and /health endpoints
│   └── schema.py        # Pydantic models, raw to T-score mapping
├── src/neuronova/
│   ├── data.py          # feature audit, target binarisation, train/val/test split
│   ├── features.py      # ColumnTransformer preprocessor
│   ├── train.py         # model training, GridSearchCV, calibration, ensemble
│   └── evaluate.py      # test metrics, fairness audit, permutation importance
├── models/
│   ├── model.joblib     # trained VotingEnsemble
│   └── meta.joblib      # model name + feature list
├── data/processed/      # train.csv, val.csv, test.csv
├── reports/             # confusion matrix, ROC/PR curves, calibration plot,
│                        # feature importance, audit report
└── web/                 # React frontend
    └── src/
        ├── components/  # Nav, ScoreQuestion, RiskResult
        └── pages/       # ScreeningForm, AuditDashboard, Methodology

---

Running locally

Create a .env file in the project root:

PYTHONPATH=src

Backend (terminal 1):

uv run --env-file .env uvicorn app.main:app --reload

Frontend (terminal 2):

cd web && npm run dev

	URL
Frontend	http://localhost:5173
Backend	http://localhost:8000
API docs	http://localhost:8000/docs

Both must be running simultaneously for the screening form to work.

---

Known limitations

• Adolescents (14+): F1 drops to 0.692. Behaviour rating scales were normed on younger children and ADHD presentation changes with age. Results for this group should be treated with additional caution. A warning is surfaced in the API response and UI.
• Dataset size: 947 participants from a research setting is small. Predictions may not generalise equally to all populations, schools, or cultural contexts.
• Self-report bias: scores depend on the accuracy and consistency of the person completing the form. The same child may score differently depending on who answers.
• Not a replacement for clinical judgement: this tool is a decision support aid. Clinical judgement, context, and the full history of a child should always take precedence.

---

Author

Built by Adit Sawhney

Disclaimer

This tool does not provide medical diagnoses. It is intended as a decision support aid to help identify children who may benefit from a formal clinical assessment. All results should be interpreted by a qualified professional.