The Creator of Apple’s FaceID Aims to Use AI for Brain Health Analysis

The co-creator of Apple’s FaceID and Vision Pro technology has dedicated the last six years to developing a cutting-edge artificial intelligence model that may eventually assist in decoding brain electrical activity for diagnosing cognitive disorders.

Recently, Gidi Littwin’s startup, Hemispheric, secured $52 million in funding after compiling data on 100,000 brain scans to train deep learning models capable of non-invasive brain examination.

Littwin departed from Apple in 2020 in search of a new direction. This shift happened when his Hemispheric cofounder, Hagai Lalazar, reached out to him via LinkedIn. Lalazar had started creating AI technology to explore the brain without surgical intervention and needed a commercially savvy cofounder to propel the business. By the time he contacted Littwin, he had already met with about 75 other candidates.

Having contributed to the development of FaceID, Littwin was then engaged in hand-tracking for an augmented reality initiative, the Vision Pro. For this project, he collected what he shared with WIRED was “hundreds of thousands of subjects’ worth of data” to train the deep learning models that powered the technology.

“Massive data collection efforts were behind these projects, and we understood we needed to create something very similar at Hemispheric,” Littwin explains, “and we have.”

Since each person’s brain activity is unique, doctors have generally depended on subjective surveys and behavioral assessments to diagnose conditions such as depression, Alzheimer’s, and Parkinson’s. To overcome this limitation, Littwin and Hagai gathered their “most valued asset”: a quarter of a million hours of brain data from 100,000 compensated volunteers in Asia, Tel Aviv, and Boston. Participants completed various activities resembling games, which stimulated different areas of their brains.

This data contributed to training a pioneering model that infers brain function from electrical activity within the skull, similar to how large language models extract meaning through statistical analysis of text. They subsequently tested this generalized model on subsets of individuals with PTSD, schizophrenia, and depression, finding that it could accurately assess their brain health. The team is currently conducting a clinical study to determine if their model can both diagnose and predict Alzheimer’s.

The team plans to submit their first product, aimed at studying PTSD, for FDA approval early next year. They hope to launch it to the public later in 2027.

To assist in diagnosing cognitive disorders, a patient dons a lightweight EEG headset that records brain electrical activity for around 15 minutes while engaging with an app on a tablet. Hemispheric asserts that its AI model will help clinicians analyze the signals for diagnoses, identify the most effective interventions by predicting treatment outcomes, and track progress.

“We envision a future where this process is as simple as a blood test,” Lalazar states. “The device will be extremely affordable and can be distributed throughout mental health clinics, hospitals, and even psychologists’ offices.”

AI-enhanced diagnostic tools for conditions such as lung cancer are already being used clinically, accelerating treatment access across Europe. Meanwhile, AI giants like OpenAI and Anthropic are venturing into healthcare, increasing competition among numerous startups in the sector.

Hemispheric has attracted early-stage investment from American and Israeli venture capital firms and private investors, including early Uber supporter Howard Morgan. The funds will be used to strengthen partnerships with governments, healthcare organizations, and pharmaceutical companies, expand their US team, and pursue regulatory approvals. Additionally, they plan to collect more brain data from millions of individuals to enhance their model.

The duo is also working on their own brain scanners to gather information they believe will yield more valuable data for their models than traditional EEGs. “These devices were never designed for machine learning, let alone deep learning,” Littwin mentions.