NeuroSense — Hirak Parekh

// THE IDEA

High accuracy isn't proof.

A score tells you how often a model is right, never why. In medicine, a model that is right for the wrong reason is worse than useless, because it looks trustworthy.

So NeuroSense treats explanations as something to test, not decoration. Each model explains itself, and those explanations are measured to see whether the model is looking at the disease or at a shortcut.

// THREE SIGNALS

Voice, handwriting, brain scans.

VOICE

A sustained "aaah"

15 acoustic features into a Random Forest, explained with TreeSHAP: which features pushed this recording towards Parkinson's, and by how much.

HANDWRITING

A spiral and a wave

A fine-tuned ResNet18 per drawing, explained with Grad-CAM heatmaps of where the network looked. Kept next to a classic HOG + SVM baseline.

IMAGING

MRI and DaT-SPECT

Split by patient, never by image, so no person's scans appear in both training and testing.

// TESTING THE EXPLANATIONS

Five numbers for every heatmap.

A heatmap can look convincing and mean nothing. I wrote five metrics once, unit-tested them on synthetic images with a known signal hidden in them, and ran every image model through the same set.

Deletion

Remove the pixels it calls important. Does confidence collapse?

Insertion

Reveal only those pixels. Does confidence come back?

On target

How much of the heat lands on the object that should matter?

Randomisation

Scramble the weights. A real explanation should fall apart.

Stability

Add faint noise. A real explanation should hold still.

// WHAT THE AUDIT CAUGHT

Three shortcuts accuracy would have hidden.

VOICE

The voice model was mostly reading age.

SHAP ranked age as the single most important input. Retraining under 25 cross-validation folds confirmed it: age alone scored as well as every acoustic feature plus age and sex combined.

Bar chart of accuracy and ROC-AUC for four voice feature sets; age alone matches the full model — **Four feature sets, 25 folds each.** Age alone (C) lands right next to all 17 features (A).

HANDWRITING

The drawing models were reading the paper.

The CNNs beat their SVM baselines and reached 0.97 AUC. The on-target metric showed why that number couldn't be trusted: over 99% of the Grad-CAM heat sat on margins, corners and paper texture, not on the pen strokes.

Insertion beat deletion every time, so the heatmaps were faithful to the model. The model just wasn't looking at the tremor. Faithful and meaningful are two different tests, which is why both are measured.

Eight Grad-CAM overlays on spiral and wave drawings, with heat on the margins rather than the strokes — **Eight drawings, all classified correctly.** Look where the heat is: the edges and corners of the page, rarely the line itself.

MRI

The MRI model was sorting scan types.

A perfect training score next to a below-chance test score was the clue. Each patient's folder mixed several kinds of scan, and the network had memorised the mix. I rebuilt the preparation to pick one structural series per patient.

Then the real answer showed up, on a second, independent dataset too: structural MRI shows anatomy, but early Parkinson's is a dopamine problem. No model can find a signal that isn't in the image. So I changed the image.

// THE FIX

Measure what a doctor looks at.

A DaT-SPECT scan shows the brain's dopamine system directly. In a healthy brain, two regions light up as commas. In Parkinson's, the tail fades and the comma shrinks into a dot.

Instead of asking a CNN to rediscover that, I measured it: five shape features, how comma-like each side is and how much the two sides differ, fed into a logistic regression. Because the model is linear, its explanation is exact: the bars add up to the score.

Three DaT-SPECT scans with outlined striatal regions and bar charts of each feature's contribution — **Three unseen patients.** Top: the detected regions. Bottom: how much each feature pushed the score. The healthy scan's crisp commas push hard towards healthy.

It held up under every check I threw at it: shuffled labels (200 times, p = 0.005), 20 alternative hand-set parameters, a control for image-export shortcuts, and added noise that flipped only 3.6% of decisions.

// RESULTS

Five features beat a fine-tuned CNN.

0.86

Balanced accuracy on patients the model never saw, over ten repeats of 5-fold cross-validation

+18 pts

Over a fine-tuned ResNet18 on the exact same folds

10×

Smaller gap between training and test scores than the CNN: 2.5 points instead of 25

Features, each one something a clinician already reads off the scan

// THE APP

One upload, an answer and its reasons.

A Flask app takes a voice clip, a drawing or a scan and returns the prediction, its probability, the heatmap or SHAP breakdown, and a plain-language note on what the model used. The classic baselines sit right next to the CNNs, so any gain is visible.

It never crashes on a bad input: if weights, a library or an explainer fail, you still get the prediction with an alert. Seven models, tested metrics, smoke-tested routes, and every number in this project regenerates from saved results.