Patient at bedside

Moving towards a
clinical world model

The Team

We've built
and shipped

1
Sohan Japa
Sohan Japa, MD MBA
Stanford ×3 · Harvard Residency
Founder ×2
Built VA San Diego Hospital-in-Home (acute care)
Shark Tank S7 (filmed)
Ben Glicksberg
Ben Glicksberg, PhD
Founding Director, AI + Children's Health Center, Mt. Sinai
h-index 66 · top 1%
PhD Mt. Sinai · Postdoc UCSF
Former VP ML, Character Bio ($93M Series B)
Advisors Dr. Girish Nadkarni Chief AI Officer · Chair, Dept of AI & Human Health, Mt. Sinai Dr. Alexander Charney Vice Chair, AI & Human Health, Mt. Sinai
2

The Problem

Clinical AI is trained on
compressed proxies

3

The Problem

The vitals and labs looked stable.

But she spoke in fragments. Her neck strained with each breath.

The doctor saw it. The algorithm couldn't.

4

The Insight

Multimodal AI needs
audiovisual information

5
Today Patient Clinical note subjective, compressed 138 Lab value periodic, delayed Imaging static, point-in-time 2 Visual scoring tools rich observation → one digit AI model trains on proxies WHAT DOCTORS OBSERVE never recorded · never modeled
6

The Solution

Digitize what
doctors observe

7

What We Measure

INDICATION MOVEMENT VOICE FACE Parkinson's lead indication Tremor Shuffling gait Hypophonia · Speech rate Articulation Masked Affect Blink rate Respiratory Distress hypoxemia · acute Tripoding · accessory muscles Paradoxical breathing Fragmented speech Breathless pauses Grimacing Air hunger Depression + neuropsychiatry Psychomotor slowing Agitation Monotone prosody Slowed speech Negative Affect Reduced expressivity
8

The Vision

A world
clinical model

Not pattern recognition - a causal model of patient state. It learns why patients decline, not just what decline looks like.

9

The Foundation

The missing piece was always the phenotype.

AV Phenotypes tremor · gait · voice · affect NEW ✦ EHR · Labs · Vitals structured clinical data Imaging · Waveforms radiology, ECG, EEG Outcomes response, progression, survival World Model causal · generative Who responds to treatment trial enrichment · responder prediction How disease progresses continuous phenotypic trajectories What drives outcomes causal links: phenotype → treatment → endpoint
10

World model. Three unlocks.

01
Pharma
High-fidelity phenotypes for trial enrichment, endpoint measurement, and RWE.
→
02
Care access
Specialist-grade assessment anywhere. Rural Montana. No travel required.
→
03
Telehealth
Visits that end with "go get imaging" can close remotely.
11

Traction

We're not
starting from zero

12
5 yrs
prior AI research
5
IRB-approved video trials
100+
patients enrolled

Traction

Spinning out of the Hasso Plattner Digital Discovery Program.

Parkinson's Disease
Cardiovascular
Neuropsychiatry
Pressure Ulcer
Pediatric Movement Disorders

13

How It Works

Edge-first.
Privacy-by-design.

14
Meta Ray-Ban glasses
Apple Vision Pro

Privacy Architecture

Capture EDGE AI on-device processing Extract FEATURES → secure cloud RAW VIDEO never transmitted
15

The Market

Disease focus
as wedge strategy

16
DRAFT
$2.5B+
Pharma PD R&D / year
200+
Active trials

The Market

UPDRS is subjective, episodic, high-variance.
Pharma needs objective, continuous biomarkers.

Better patient selection · Smaller trials · Faster readouts

UPDRS = Unified Parkinson's Disease Rating Scale

17
DRAFT

Business Model

We charge pharma for what they already budget:
better trial data.

pre-trial during trial post-market PATIENT STRATIFICATION Per screen Identify likely responders before trial Near-term revenue DIGITAL ENDPOINTS Milestone AV capture as regulatory endpoint High-value contract DATA LICENSING Annual license Proprietary AV dataset drug discovery + RWE Long-term · recurring
18

Landscape

Multi-modal fusion (video + audio + EHR + outcomes) enables causal discovery.

NOVEL + GENERATIVE Existing data modalities Novel observational data Descriptive Generative EHR / NLP AI Flatiron · Komodo Digital Twins Unlearn · Q Bio Wearables Apple · Empatica Point solutions Rune Labs Altoida Newco
19

Roadmap

Where we're
going

20

Roadmap

Building the data moat to train a world model.

NOW YEAR 1 YEAR 2 Q1 Team assembled 5 sites activated Q2–Q3 500 patients enrolled PD + 2nd indication 1,000 PATIENTS First pharma pilot 10+ sites Q5–Q7 International expansion 3 indications · LOIs 10,000+ PATIENTS World model threshold Pharma revenue 10× data scaling
21
Patient at bedside

Thank you

Appendix · Competitive Detail

Point solutions pick one signal for one disease.
We're building the generalizable layer.

MULTI- MODAL PASSIVE CAPTURE NO WEARABLE DISEASE- AGNOSTIC CAUSAL MODEL WHAT THEY MEASURE Newco us ✓ ✓ ✓ ✓ ✓ Movement · Voice · Face All diseases Sonde Health voice · mental health ✗ ~ ✓ ~ ✗ Voice only → depression / respiratory Rune Labs Apple Watch · Parkinson's ✗ ✓ ✗ ✗ ✗ Apple Watch motion → PD tremor only Altoida AR tasks · Alzheimer's ~ ✗ ✓ ✗ ✗ Active AR task → Alzheimer's prediction ✓ = yes ✗ = no ~ = partial
23

Appendix · Competitive Funding

Point solutions are well-funded — and still single-signal.

FOUNDED RAISED NOTABLE Sonde Health voice biomarkers mental health + respiratory 2016 $54.5M Series B $19M (2022) Voice-only · Air Force contract 2024 Rune Labs Apple Watch motion · Parkinson's 2018 $52M+ $11.2M (Jun 2025) Wearable required · PD only · FDA digital endpoint Altoida AR tasks · Alzheimer's prediction 2014 $24M+ Merck KGaA-led (Feb 2024) FDA Breakthrough · Alzheimer's only · active task Funding per Crunchbase
24

Appendix · The Gap

Same patient. Different signals.

Today Patient EHR Data Vitals ✓ Labs ✓ Imaging ✓ Notes ✓ Clinical AI structured data only ✓ STABLE With AV Patient + Video EHR Data Vitals Labs Imaging Notes + AV Phenotypes real-time · dynamic Breathing Posture Face Speech World Model multi-modal fusion ⚠ DETERIORATING
25