Applied AI for personal healthcare

In order to deliver on our mission, we're researching at the intersection of healthcare & AI.

Our research focuses on three domains.

Health record synthesis

First, we endeavor to solve one of healthcare’s most challenging problems: interoperability. Doctors control your health data, so your medical context is scattered across various health portals. Regulation requires data interoperability, but in practice, health data remains siloed and only aggregated on an as-needed basis. Thankfully, every individual has a right to their health data, and doctors must comply with HIPAA release of information requests.

Our first research objective is using AI to reliably obtain a user’s health records (across multiple specialties), combine them with device data (iPhone & Apple Watch), and coherently synthesize them into one unified AI-ready health record. This process currently requires an enormous amount of personal effort: emails, phone calls, digging for information across the internet, your inbox, and your calendar. We imagine a world where, with minimal patient information, Valta can perform this entire process. Moreover, we seek to sustain a unified health record, where new visits to a doctor are automatically picked up via a calendar integration, and data requests happen in the background.

Fundamentally, this problem is composed of:

Inbox & calendar health sync: accurately extract healthcare events from your calendar and healthcare communications from your inbox
Fuzzy provider search: find a specific healthcare provider & facility given partial or imperfect information
Reliable health data requests: given a user and a provider or insurance plan, reliably retrieve health records via email, phone, or API
Health record parsing and organization: parse records across time & specialties into a navigable & useful knowledge tree

You shouldn’t have to work for your health record – your health record should work for you.

Clinical effectiveness

Delivering a superhuman health assistant is an extremely tall task. Each use case and medical specialty poses its own research & product challenges. Consider dermatology: Users should be able to “scan” their skin, and Valta (with context of their skin history and previous biopsies), should give them a risk assessment (diagnosis), a recommendation of what to do next (treatment planning), and help them schedule a biopsy if needed (care coordination).

Across all common specialties, Valta needs to be able to:

Diagnose: observations → probability-weighted diagnosis
Plan treatment: diagnosis → clinically-valid treatment plan
Coordinate care: diagnosis + treatment plan → system integration + patient advocacy

In order to achieve these capabilities, we must:

Develop evaluations: define success across various clinical domains & use cases
Benchmark frontier models: measure how effective each model is
Publish research: share results, request critiques, and iterate with the scientific community

If we make useful clinical competencies functionally verifiable, AI models can hill climb against clear goals and eventually achieve superhuman clinical performance.

Actuary outcomes

Many individuals expect their employer or the government to underwrite their healthcare. In order to make AI-first care a reality, Valta must transition from a business model reliant on individual early adopters to institutional healthcare customers. Risk-bearing healthcare companies (self-insured employers and payers) adopt new technology based on return on investment.

Therefore, Valta must research how equipping people with a doctor in their pocket leads to better health outcomes and prevents avoidable spending. We believe AI personal healthcare can demonstrably bend the cost curve. To unlock the potential of AI for personal healthcare, we must prove this with verifiable research & commercial case studies.

Feasibility

This sounds crazy; will it actually work? We think so.

Technically, we believe this is achievable. In early 2024, Google demonstrated with AMIE that AI can more accurately diagnose conditions via chat than human doctors.

Reasoning models, like OpenAI's o1 model, have shown rapid & sustained improvement in functionally-verifiable domains

Societally, this is highly disruptive and will likely have loud opponents. Licensed medical providers may attack the credibility of AI care. Entrenched incumbents may lobby to mandate a human in the loop. Consumers may be skeptical and fearful of clinical accuracy.

But our bet is that humans will ultimately prefer AI care: unlimited, accurate, private, and affordable. What do you think?

Join us

If you believe in this future, join the pilot and become one of the first to experience Valta.

If this sounds like your life's work, send us an email with your resume and a project you're especially proud of.