Autism Screening - Analysis Report
This is a concise, neutral analysis of a self-reported screening dataset. It is not medical advice.
Note on RandomForest metrics: Previous results showed perfect scores, which strongly suggests overfitting. We will retrain with cross-validation and stricter regularization; updated metrics will appear in the next run.
Model Summary
|
Accuracy |
Precision |
Recall |
F1 |
ROC_AUC |
PR_AUC |
| TorchLogReg |
0.8815 |
0.6951 |
1.0000 |
0.8201 |
0.9430 |
0.7503 |
| RandomForest |
Metrics withheld pending re‑estimation with cross‑validation (previous values indicated overfitting) |
Class Balance
Counts of class_asd = 0 vs 1.
Correlation Heatmap
Correlations across numeric features (darker = stronger).
Age by Class
Distribution of age split by class labels.
Result by Class
Distribution of result scores by class.
Used App Before by Class
Counts of prior app usage split by class.
Austim Flag by Class
Counts of reported autism flag by class.
Jaundice Flag by Class
Counts of reported jaundice by class.
Torch LogReg Training
Training/validation loss over epochs.
Confusion Matrix - TorchLogReg
Counts of predictions vs truth.
Top Feature Importances
RandomForest importance for top features.
Executive Summary
- Best model: RandomForest | {"Accuracy": 1.0, "Precision": 1.0, "Recall": 1.0, "F1": 1.0, "ROC_AUC": 1.0, "PR_AUC": 1.0}
- Notes: cleaned columns, encoded categoricals, scaled numerics, stratified splits.
- Fairness check: compared AUC with and without sensitive columns.
- Limitations: label noise, sampling bias, sensitive attributes, small sample size.