Wearables have made physiological data abundant, but much of the body’s chemistry and internal function remains difficult to observe. Three problems deserve sustained attention: understanding multimodal physiology, continuously measuring molecules, and monitoring organs beyond the clinic.
1. Can Foundation Models Understand Human Physiology?
A higher resting heart rate can accompany exercise, infection, alcohol consumption, or a normal menstrual-cycle transition. Interpreting it requires context across signals and time. The high-value problem is whether a model can integrate physiological measurements, behavior, personal history, and language into useful, individualized health reasoning.
Google’s research offers a concrete example. Its Personal Health Large Language Model, published in Nature Medicine in 2025, adapted Gemini to sleep and fitness applications using aggregated wearable data. Evaluations included expert knowledge, personalized recommendations, and self-reported sleep-quality prediction. These results establish capabilities on defined tasks; they do not establish better long-term health outcomes.
Two complementary directions extend this work. SensorLM aligns multimodal sensor data with language, supporting activity recognition, retrieval, and description generation. SensorFM, introduced in 2026, learns reusable representations from large-scale, unlabeled wearable data and evaluates transfer across 35 health-prediction tasks. Together, these projects suggest an architecture combining sensor representations with language-based reasoning, rather than assuming a general chatbot can directly interpret physiology.
The central challenges are generalization and validation. Models must handle missing data, device differences, changing personal baselines, and populations poorly represented in training. They must distinguish plausible explanations from supported conclusions. Prospective studies should test whether recommendations improve decisions and outcomes, beyond benchmark performance.
Xin Liu is a key researcher to follow for this intersection of sensor foundation models and personal health AI. The University of Washington’s UbiComp Lab, led by Shwetak Patel, is an important academic anchor: Liu’s doctoral work there combined physiological sensing and machine learning. The Google projects are broader team efforts. The milestone worth watching is reliable transfer to new people and health tasks with limited additional labeling.
2. Can We Continuously Measure Molecules Beyond Glucose?
Heart rate and temperature reveal physiological responses, but usually leave their molecular drivers unmeasured. Continuous hormone sensing could help connect cycle changes, fertility, stress physiology, and treatment response to measurable biochemical dynamics. The broader opportunity includes metabolites and therapeutic drugs.
Two approaches deserve attention: collecting sweat through skin-mounted microfluidics, and accessing interstitial fluid through minimally invasive interfaces. Neither makes every target equally accessible. Sampling availability, concentration, and the relationship between the sampled fluid and blood must be established for each molecule.
Wei Gao’s Caltech group provides a strong example. Its wearable estradiol sensor integrated sweat induction, microfluidic sampling, electrochemical detection, and wireless communication. Human experiments reported menstrual-cycle fluctuations and correlation between sweat and blood estradiol. The group’s 2026 publications include multiplex ovarian-hormone monitoring, extending the research toward multiple targets. These demonstrations should be distinguished from validated, multi-day consumer monitoring.
Kevin Plaxco’s UCSB laboratory addresses a complementary enabling technology: electrochemical aptamer-based sensors. These use molecular recognition to transduce target binding into electrical signals. His group has demonstrated rapid molecular measurements and feedback-controlled drug delivery in animal studies, illustrating possibilities beyond glucose-specific enzyme chemistry.
The difficult step is maintaining trustworthy measurements on the body. Low concentrations, interfering compounds, biofouling, recognition-element degradation, sampling changes, and calibration drift can undermine an otherwise sensitive sensor. Detecting a molecule once also differs from reversibly tracking rises and falls over days. Temporal lag and fluid-to-blood relationships require longitudinal validation.
The decisive milestone is a wearable that repeatedly agrees with reference assays during ordinary life, remains stable for a useful wear period, and supports a defined decision. For women’s health, that could mean demonstrating added value for cycle interpretation or treatment monitoring. More frequent measurements become valuable when their meaning and practical use are established.
3. Can Wearable Ultrasound Be Applicable Outside the Clinic?
Most consumer wearables observe accessible signals at the skin. Ultrasound can interrogate deeper structures, including blood vessels and the heart. The opportunity is to observe internal function over time, capturing changes that a brief appointment may miss.
Sheng Xu, now at Stanford, is a central researcher in this field, following extensive work at UC San Diego. A 2026 collaborative study demonstrated an abdominal ultrasound patch for continuous fetal and umbilical blood-flow monitoring, with early validation in pregnant patients. The reported system remained connected to a computer by cable; wireless home monitoring was a development goal. It illustrates both the value of continuous internal sensing and the remaining integration work.
Two MIT groups offer complementary approaches. Xuanhe Zhao’s team demonstrated bioadhesive ultrasound stickers that imaged internal structures for up to 48 hours, addressing sustained skin attachment and acoustic coupling. Canan Dağdeviren’s group develops conformable ultrasound for accessible imaging; its 2026 portable breast-imaging system reduced processing hardware substantially, while a smaller home wearable remained a future objective.
The bottleneck is the complete system. A practical device must maintain acoustic contact, find and track the target as the wearer moves, manage power and heat, and interpret data without repeated expert adjustment. Long-duration use also requires evidence on skin tolerability and acoustic exposure. An imaging prototype alone does not establish autonomous monitoring.
The most useful milestones will be application-specific: sustained blood-flow tracking, reproducible cardiac measurements, or dependable organ-volume estimates under realistic conditions. Clinical studies must then establish whether this additional information changes care. The long-term prize is access to internal physiology when and where decisions are made, with sufficiently little operational burden to make repeated monitoring practical.