
AIdeas Finalist: NeuroVoice
Small moments, repeated daily, reveal what single snapshots never can - patterns that quietly signal neurological risk. NeuroVoice is a multimodal Parkinson’s screening platform that tracks pattern over time to detect subtle neurological changes early.
Team Members: Shambhvi Sharma @shambhvi2006, @yash4428
App Category: Social Impact
Our Vision:
What if neurological decline could be detected before it becomes visible?
Many neurological diseases, including Parkinson’s, start subtly. Changes in speech and facial movements or daily behavior often appear long before a medical diagnosis can be made. Yet these signals go completely untracked between clinical visits.
NeuroVoice is here to change that.
Designed as an AI companion available through WhatsApp or a native mobile app, NeuroVoice turns a 5-minute daily interaction into structured, longitudinal biomarker data. Through natural conversation, the app analyses vocal biomarkers - jitter, shimmer, pitch variation, and speech rate, alongside facial signals including rigidity, asymmetry, and blink rate, building a trend picture over time displayed as easy-to-read progression graphs.
For individuals without a diagnosis, NeuroVoice encourages early awareness and timely medical conversations. For diagnosed patients, it provides structured progression tracking between appointments for caregivers and doctors.
NeuroVoice does not diagnose - it flags trends for clinical follow-up.
Our vision is to shift neurological care from occasional and episodic to continuous awareness integrated into everyday life. Not to replace doctors, but to make the time between clinical visits medically meaningful.
Beyond voice and facial signals, NeuroVoice extends to motor assessments, sleep tracking, and medication response, building a complete longitudinal picture no single biomarker could provide alone.
Why this matters?
Neurological diseases progress silently, making early detection extremely challenging.
By the time visible symptoms show up, window for early intervention and prevention is lost. Even though many families notice these changes, a slight tremor, slower speech, reduced facial expression, but they dismiss them as natural aging because no simple tool exists to tell them whether those changes warrant attention.
At the same time, for people already diagnosed with Parkinson's, the gap between clinic visits can stretch weeks or months. During that time, there is no structured way to know whether symptoms are stabilizing or progressing. Medication adjustments get made on memory and intuition rather than data.
In regions like India where awareness is low, consistent follow-up is limited, but even in advanced healthcare systems, continuous monitoring is still uncommon.
NeuroVoice addresses this by bringing neurological monitoring into our everyday life, for individuals this enables awareness, for caregivers it provides reassurance and for doctors it gives a way to track patterns of all their patients longitudinally between visits.
Accessibility is central to this. WhatsApp integration is a way to meet elderly users and caregivers on a platform they already use and trust, reaching populations that existing neurological monitoring tools were never designed to serve.
How We Built This?

1. Voice & Facial Analysis
NeuroVoice begins when a user sends a voice message through WhatsApp or the native app received via Amazon API Gateway, processed through AWS Lambda, authenticated via Cognito, and stored in S3, with longitudinal history in DynamoDB. Amazon Transcribe converts audio to text while Praat extracts jitter, shimmer, speech rate, pitch variation, and tremor markers. A model trained on the Parkinson's Speech with Multiple Types of Sound Recordings dataset is deployed via SageMaker, generating daily biomarker trend scores. A second interaction captures a facial recording analysed for rigidity, asymmetry, and blink rate.
On held-out test data, the model achieves ROC-AUC 0.65 on voice alone and 0.96 with UPDRS indicators. Both figures come from clean, lab-recorded audio real-world validation through a clinical pilot is our immediate next step.
2. Privacy & Security
NeuroVoice runs a deliberate two-tier privacy model. WhatsApp is our accessibility-first channel for elderly users, caregivers, and rural populations across India consent-based, voice and facial only, aligned with India's DPDP Act 2023. HIPAA does not apply to our primary market.
The native app is the clinical-grade tier end-to-end encrypted, DPDP and HIPAA-aligned, with strict data minimization and explicit consent flows. For clinical partnerships or US deployment, the native app is the compliant path. WhatsApp handles accessibility. The native app handles depth.
3. Motor, Sleep & Medication Assessments
The native app expands NeuroVoice into a clinically robust platform. Four motor tests - finger tap, Archimedes spiral, meander wave, and clock drawing are processed via Lambda for real-time UPDRS-aligned scoring, with results stored in DynamoDB. For spiral and clock drawing, a model trained on the Parkinson's Disease Spiral Drawings Using Digitized Graphics dataset is deployed via SageMaker, giving motor assessment a validated research foundation.
Sleep monitoring covers RBD detection, REM tracking, resting heart rate, and dystonia logging via SageMaker endpoints. The medication tracker runs event-driven on EventBridge and Lambda logging adherence, detecting on/off periods, and triggering caregiver alerts via SNS. The support network groups patients by Hoehn and Yahr stage using DynamoDB metadata.
All signals - voice, facial, motor, sleep, and medication, consolidate through a Step Functions pipeline into a single continuously evolving longitudinal profile.
4. Conversational Layer & Dashboards
Amazon Bedrock powers the conversational layer, capturing sleep quality, mood, motor difficulty, and sensory context across every interaction. Amazon Polly returns responses in natural voice.
Caregiver and doctor dashboards are built on AWS Amplify, with automated monthly summaries via EventBridge and SES.
5. Rehabilitation
The whole-body movement test replicates standard neurological walking evaluations 8 to 10 steps forward, turn, walk back while the system observes arm swing symmetry, stride length, posture, and turning coordination. Detection focuses on deviation from the user's own baseline, because natural movement reveals what structured tasks miss.
The gamified LSVT BIG exercise guides users through large-amplitude reaches in multiple directions, with MediaPipe providing real-time pose feedback on extension and whole-body engagement. Repeated consistently, LSVT BIG directly retrains motor output -improving gait, posture, and daily functional movement.
In production, pose estimation will run via SageMaker with Step Functions orchestration. Rehabilitation outputs feed directly into the multimodal biomarker pipeline - therapy that is not just therapeutic but measurable.
Demo:
Voice & Facial Analysis:

The user completes a guided voice recording, sustained sounds and a short passage, while Praat extracts jitter, shimmer, pitch variation, and speech rate in real time. A 30-second facial recording follows, with MediaPipe mapping blink rate, asymmetry, and rigidity. Results land as an overall risk score with full biomarker breakdowns, tracked historically over time. On WhatsApp, the same voice pipeline runs without any app download - a voice note in, a risk score out.

Motor Assessments:

Four structured tests evaluate motor function. The Archimedes Spiral and Meander Wave measure tremor, deviation, and speed consistency against UPDRS benchmarks, clinical findings and full biomarker breakdowns are immediate. The Finger Tapping Test scores each hand separately over 10 seconds, measuring tap frequency, hesitations, and fatigue ratio to MDS-UPDRS Part III standards.

Sleep, Medication & Support:
Sleep & Physiology flags RBD indicators, tracks REM fragmentation, deep sleep, and resting heart rate nightly. The Medication Tracker logs doses, maps a daily on/off timeline, identifies peak off-period windows, and triggers caregiver alerts automatically. The Support Network connects users to a Hoehn and Yahr stage-matched peer group, a live chat where patients at the same progression stage support each other daily.
Rehabilitation:
The walking test observes natural gait, arm swing symmetry, stride length, posture, and turning coordination detecting deviation from the user's own baseline over time. The gamified LSVT BIG exercise uses MediaPipe pose tracking to give real-time feedback on movement amplitude, guiding users through large-reach exercises that directly retrain motor output. Therapy that is measurable, not just instructional.
What We Learned?
The feedback that resonated most was the question about clinical validation not because it was unexpected, but because it forced us to be precise about what we were actually claiming. A 0.65 ROC score is honest, but honesty without context isn't enough. The real answer is longitudinal data a single interaction says little, but trends tell the real story. That reframe changed how we positioned everything. NeuroVoice flags trends for clinical follow-up, it does not diagnose. That is not a disclaimer, it is the product. What came out the other side of that interrogation was a cleaner, more defensible version of the idea.
The deepest insight wasn't about the model. It was about the gap. Neurological decline happens in the months between appointments, in daily conversation, in patterns no quarterly assessment ever captures. The infrastructure to close that gap conversational AI, multimodal signals, longitudinal tracking already exists. What was missing was something that brought it together in a form that people would actually use. That is what NeuroVoice is trying to be.
Competitive Landscape:
App | What it does | NeuroVoice's Difference |
|---|---|---|
Khayyal/ Speak PD | Voice-only Parkinson's screening | NeuroVoice adds facial + behavioral context |
Roche / Floodlight One | Sensor + app-based Parkinson's monitoring | Requires dedicated hardware/setup; not WhatsApp-accessible |
mPower (Sage Bionetworks) | Research-focused phone sensor data collection | Research tool, not a user-facing wellness companion |
Winterlight Labs | Clinical voice analysis for cognitive decline | Enterprise/clinical only, not consumer accessible |
Business Model
The primary channel is B2B2C - clinics and neurology practices offer NeuroVoice to patients as a monitoring layer between visits, with longitudinal biomarker data flowing back into clinical review.
But the model doesn't stop at the clinic door. A direct consumer tier targets individuals over 40 who want to track their own cognitive health proactively, not because something is wrong, but because catching a trend early, before it becomes a diagnosis, is exactly the kind of agency people are increasingly willing to pay for.
Both channels feed the same underlying data asset: anonymized, data, licensed to pharmaceutical and research partners as a secondary revenue stream. Federated learning sits underneath the entire model, the system improves with use without raw data ever leaving the device. Clinical adoption and consumer uptake compound each other. The data asset grows with both.
Responsible healthcare technology must be transparent about what it cannot do. NeuroVoice is not a diagnostic tool - it is the infrastructure for making the time between clinical visits medically meaningful.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article