5 Early-Stage AI Health Startups Built on Clinical Data

Why CT Provision stands out

CT Provision is built around data already generated at hospital admission: brain CT plus clinical variables. Its HOP model has been retrospectively evaluated on 127 patients with subarachnoid haemorrhage, with reported AUCs of 0.731 for vasospasm or delayed cerebral ischaemia and 0.857 for death.

That makes the AI claim concrete. The model is not searching for a generic imaging pattern; it is designed to estimate defined complications after a specific neurological emergency.

What CT Provision is building

HOP derives quantitative features from CT images and combines them with variables such as age, sex and neurological severity scores to generate patient-level risk estimates.

Why OneCareAI stands out

OneCareAI uses an established cardiovascular signal rather than requiring a new test. Its GALEN technology originated at the Barcelona Supercomputing Center, with validation and pilots running at hospitals in Spain and Italy.

The product thesis is that ECG contains information relevant to stroke risk that can be extracted computationally and combined with other clinical context.

What OneCareAI is building

INVICTUS applies GALEN to a single ECG and can combine the trace with genetic, molecular, clinical and lifestyle data to estimate stroke risk.

The practical advantage is compatibility with conventional ECG machines and wearables. For the AI to create value, the model has to turn an existing signal into risk information that can be acted on without adding a separate diagnostic pathway.

Why Oculum stands out

Oculum applies AI to live surgical imaging rather than to an offline diagnostic dataset. The startup is running a second clinical study at Sunnybrook and is developing a dye-free way to visualize nearly transparent retinal tissue during surgery.

The data-to-decision link is immediate: the model changes what the surgeon can see while the procedure is happening.

What Oculum is building

Oculum is developing a modular add-on for surgical microscopes that applies AI-powered virtual staining to the live image.

Instead of asking surgeons to interpret a separate software output, the product aims to place the enhanced tissue information inside the view they already use. That makes latency, image fidelity and workflow integration as important as model performance.

Why Mana Medical stands out

Mana Medical combines wearable data with short symptom check-ins and uses the output for two decisions: daily pacing guidance for patients and prioritization for clinicians. The startup has received pre-seed backing from Zinc and is running an eight-week beta study with recruitment support from the National Rheumatoid Arthritis Society.

The AI therefore sits between continuous patient-generated data and a care workflow rather than operating as a standalone prediction tool.

What Mana Medical is building

Patients connect an existing wearable and record symptoms. Machine learning combines the signals into personalized activity guidance, while a clinician dashboard calculates a Patient Stability Index and ranks patients who may need attention.

The commercial question is whether those data can make chronic-care follow-up more proactive. If the model identifies meaningful changes between appointments, the value comes from changing who acts and when, not from collecting more wearable data.

Why I See U Baby Care stands out

I See U Baby Care turns video from neonatal care into sleep-wake information without adding contact sensors. Founder Chanel Sam co-authored research on the BeSSPI sleep-classification method in 69 premature infants, covering more than 10,000 minutes of observation.

The startup has also secured a Dutch hospital collaboration, giving the technology a route to be developed in the environment where the data are actually produced.

What I See U Baby Care is building

The system uses a camera and machine-learning software to classify whether a premature baby is asleep or awake and present that information to care teams in real time.

That connects a passive data stream to a concrete neonatal-care objective: protecting sleep while reducing additional sensors on vulnerable infants. The value of the AI depends on whether the classification remains reliable enough to influence care routines in a busy clinical setting.