Parkinson's Disease Detection

About the Parkinson's disease detection M.Tech project

Parkinson's disease often shows early signs in two places: the voice turns breathy and shaky, and handwriting gets small and trembling. A smartphone can capture both at home. This Parkinson's disease detection M.Tech project asks the user to hold an "aaa" sound for three seconds and to trace a spiral with a finger, then combines the two results into one screening score.

The research question: can two calibrated phone tests, fused late, match or beat a heavier attention-based fusion, and still give a sensible score when the user skips one test?

Research behind it

Base paper. The work builds on Parkinson's Disease Prediction: An Attention-Based Multimodal Fusion Framework Using Handwriting and Clinical Data, published in Diagnostics (MDPI, 2025). It passes handwriting images, spiral images and clinical records through self-attention and cross-attention, joins ResNet and hand-crafted features, and decides with an SVM.

Research gap. The framework does not use voice, which is one of the earliest and easiest signs to record. Its drawings are made on paper and scanned, so speed and timing are lost. Feature-level fusion also needs every test from the same person, and public voice and drawing datasets come from different people.

Proposed system and novelty. Voice is described by jitter, shimmer, harmonics-to-noise ratio and MFCC statistics for XGBoost. The spiral is read twice: as an image by MobileNetV3 and as a touch trace that gives speed variation and radial deviation. Each output is calibrated, and a learned weighted sum of log-odds fuses whatever tests are available. Subject-wise folds keep all recordings of one person on the same side of the split.

Models compared. Voice-only SVM and XGBoost, spiral-only ResNet50, simple averaging and the base paper attention fusion re-implemented, against the proposed calibrated fusion. The ablation removes calibration and trace features, uses equal weights and compares record-wise with subject-wise splits.

If you would like another phone-sensor health topic, see the elderly fall detection project, which uses accelerometer data with a CNN-LSTM.

How it works

• The user records a steady "aaa" sound while a meter checks the volume.
• The user traces a spiral over a faint guide with one finger.
• The app uploads the audio, the image and the touch trace to the Flask API.
• Voice features, the spiral image and trace dynamics are scored separately.
• Calibrated scores are fused, even if one test was skipped.
• The app shows each score, the fused result and advice to see a neurologist.
• The user can share the test, and the neurologist adds remarks.

Users and modules

There are three roles. The user takes the two tests, sees the scores and history and downloads a summary. The neurologist reviews shared tests, plays the audio, replays the drawing and sends remarks. The admin manages users and records, checks the dashboard and model comparison page, and sets the fusion weights and alert limit. All models are small and run on a laptop CPU.

What you get

• Phase 1 Research: Base paper study: attention-based multimodal fusion for Parkinson's (Diagnostics, MDPI, 2025); Literature survey of 15-20 recent Parkinson's detection papers; Research gap, novelty and objectives; Voice and spiral dataset EDA notebooks; Synopsis and Phase-1 report.
• Phase 2 Implementation: Android app (Java) with voice and drawing tests; Flask API, neurologist and admin panels; Voice features with XGBoost model; MobileNetV3 spiral model and trace features; Calibrated late fusion and base fusion re-implementation; MySQL database and trained models.
• Results: Sensitivity, specificity and ROC-AUC per test; Fusion vs single test table, missing-test results; Subject-wise vs record-wise split comparison; results.xlsx and all graphs in high resolution.
• Thesis and paper: Thesis written for you (about 90 pages); IEEE conference paper draft; Journal paper draft; Publication guide.
• Reviews, viva and support: PPTs for Review 1, Review 2 and final viva; Viva book with 50+ questions; A1 poster and project diary; Demo and code walkthrough videos; 1-on-1 setup, 2 revisions and support till viva.

Who can use this project

M.Tech and M.E. students of CSE and IT who want a multimodal healthcare dissertation can use this as a research base. It matches searches for a multimodal fusion voice and spiral drawing project, a calibrated late fusion healthcare project or a voice feature XGBoost CNN project.