What do xMAE and HiMAE actually do?

xMAE combines two heart signals from a wearable — ECG and PPG — to estimate continuous cardiac data without a manual ECG reading, while HiMAE analyzes short- and long-term wearable data patterns, like individual heartbeats and sleep cycles, together on the device.

Is my health data processed on the device or sent to the cloud?

Samsung says HiMAE runs its analysis directly on a smartwatch-class processor in under one millisecond, meaning the raw signal processing happens on-device rather than on a server.

When will xMAE and HiMAE be available on a Galaxy Watch?

Samsung has not announced a launch date or confirmed the models will ship in a specific product — so far it has only published the research and peer-reviewed papers.

How is this different from OpenAI's ChatGPT Health?

ChatGPT Health is a cloud-based chatbot feature for health questions, while xMAE and HiMAE are on-device models that interpret raw sensor signals from wearables, like heart rate and sleep data — a different layer of the health AI stack.

Samsung Unveils xMAE and HiMAE AI Models for Wearables

Samsung Research introduced two new AI models built to make sense of the raw biosignals that wearable devices already collect, the company said in an August 14 announcement. The models, called xMAE and HiMAE, are designed to run directly on a smartwatch rather than sending sensor data to the cloud, part of a broader push toward on-device AI.

Reading two signals at once

The first model, xMAE, learns the timing relationship between two heart-related signals that wearables already record: electrocardiogram (ECG) readings and photoplethysmography (PPG), the light-based pulse sensor built into most smartwatches. Samsung compared the technique to “measuring the delay between seeing lightning and hearing thunder” — by learning that relationship, xMAE can estimate continuous cardiac information from PPG data alone, without a person taking a manual ECG reading. In Samsung’s own tests, the model outperformed both single-signal models and earlier multimodal approaches on 15 of 19 evaluation tasks, including cardiovascular disease prediction, abnormal test result detection and sleep-stage classification.

The second model, HiMAE, addresses a different problem: wearable data carries useful patterns at very different time scales, from a single heartbeat lasting a fraction of a second to a sleep cycle spanning hours. Samsung says HiMAE processes short and long windows of data together and can run its analysis on a smartwatch-class processor in under one millisecond, which the company describes as the first on-device health foundation model to do so in real time.

Peer-reviewed, but no launch date

Samsung Research America’s Digital Health Team detailed the two models in papers accepted at separate machine-learning venues: xMAE at the International Conference on Machine Learning (ICML) and HiMAE at the International Conference on Learning Representations (ICLR). Samsung frames both as a step toward its “Connected Care” vision, outlined at its Galaxy Unpacked health forum in July, in which wearables continuously watch for meaningful health changes instead of waiting for a person to notice symptoms. The company has not said when, or whether, xMAE and HiMAE will ship inside a Galaxy Watch product.

The research adds to a wider move by device and AI makers into consumer health tools: OpenAI opened its ChatGPT Health feature to all US adults earlier this month, and other efficient, locally run models — such as PrismML’s Bonsai 27B, a large language model adapted to run on a phone — point to the same trend of pushing AI processing onto personal devices.

Reading two signals at once

Peer-reviewed, but no launch date

Frequently asked questions