Healthleap raises $38M to turn hospital notes into patient risk signals

Healthleap has raised $38 million to expand an AI platform that screens hospital records for patients who may need additional clinical attention. Language models extract evidence from clinicians’ notes, which risk models then combine with lab results, vital signs, and other structured data.

The funding, reported by TechCrunch on October 7, comprises an $8 million seed round co-led by Sequoia Capital and First Round Capital and a $30 million Series A led by Hummingbird Ventures. Healthleap hasn’t disclosed its valuation.

The company says its software runs in more than 50 hospitals, up from three hospital partners a year ago. Customers include Penn Medicine, Cedars-Sinai, Intermountain, Houston Methodist, and Emory Healthcare. Revenue has grown more than tenfold over that period, according to CEO Josiah Meyer, though he hasn’t provided an absolute figure.

That’s substantial growth in hospital adoption. Judging the screening system itself is harder: TechCrunch’s report includes financial impact claims but no sensitivity, specificity, or prospective clinical outcome results.

Finding evidence in clinical notes

Founded in South Africa in 2022 by siblings Jemima and Josiah Meyer, Healthleap began with a clinical nutrition tool Jemima built for dietitians. It later expanded into screening for overlooked conditions, including malnutrition and delirium.

Malnutrition is a sensible starting point. Research estimates that it affects 20% to 50% of hospital inpatients, depending on the population and assessment method. It’s associated with longer hospital stays, impaired wound healing, infections, and other complications.

Some relevant evidence sits in database fields: weights, laboratory measurements, medications, and diet orders. Other evidence appears only in narrative documentation. A clinician might record poor appetite, difficulty swallowing, recent weight loss, or muscle loss without entering each observation as a discrete field.

Healthleap uses language models to extract those clinical concepts, including whether a note affirms or negates them. Its risk models combine the extracted information with structured patient data.

This gives the language model a bounded extraction task. In principle, engineers can evaluate extraction errors separately from errors in the downstream risk score. The report doesn’t specify which models Healthleap uses, how it trains them, or how it implements that separation.

Negation is one of several extraction problems. “Patient denies difficulty swallowing” and “patient reports difficulty swallowing” should produce different signals, as should a current observation and a symptom documented six months earlier. Copied notes, conflicting assessments, and references to someone other than the patient complicate the job further.

These are familiar clinical NLP problems. Syntactically valid output can still misrepresent a record, with consequences for which patients get reviewed.

Nightly screening and its limits

Each night, Healthleap analyzes every adult inpatient’s record, according to Meyer. Inputs include labs, vitals, weights, medications, diet orders, diagnoses, and clinical notes. Each morning, it writes a risk score into the care team’s existing workflow. A dashboard provides additional patient trends.

The company says the software flags patients for review; it doesn’t diagnose them.

Nightly processing gives systematic screening a predictable processing window. Staff don’t have to open a chatbot or assemble records themselves, and delivering results through an existing workflow reduces the friction that can leave a competent hospital tool unused.

The cadence also means a score can lag behind a patient’s changing condition. Whether that delay is acceptable depends on the screening task and how clinicians use the result. Hospitals shouldn’t treat a batch screening system as real-time monitoring.

Deployment across 50-plus hospitals brings problems beyond model throughput. Documentation conventions, units, missing-data patterns, and EHR configurations vary. An extraction pipeline that works well with one institution’s notes may behave differently when another uses unfamiliar abbreviations or heavily templated documentation.

Technical buyers have specific questions to ask: How are amended notes handled? What happens when a feed arrives late? Can the team trace a score back to the source observations and the model version that processed them?

Standards such as HL7 FHIR can help exchange clinical resources, but local differences in meaning and data quality remain. The report doesn’t disclose Healthleap’s integration protocols.

How much work do the flags create?

Healthleap produces a list of patients for clinicians to review. The time and staff available for that review affect how useful the system can be.

A low threshold may catch more true cases while sending more false positives to clinicians. A higher threshold can reduce workload but miss patients who would benefit from assessment. Hospitals need to know how a threshold behaves in their population and what staffing it assumes.

Useful evaluation would include precision at the hospital’s actual review capacity, missed-case rates, calibration, and subgroup performance. A well-calibrated score should correspond reasonably to observed risk. Performance also needs checking across patient groups, specialties, and institutions; one aggregate number won’t cover those differences.

TechCrunch’s report doesn’t provide those measurements, so readers can’t judge how consistently Healthleap finds cases or how much additional work its flags create.

Clinicians also need enough supporting evidence to check why a patient surfaced, especially when the score relies on text extraction. Source-note references, timestamps, and visible distinctions between current and historical observations would help. Buyers should ask for those capabilities; the report doesn’t confirm that Healthleap provides them.

Security needs the same scrutiny. Processing broad inpatient records brings sensitive notes and other protected health information into the pipeline. Buyers need specifics on inference hosting, retention, access controls, audit logs, and whether external model providers receive patient data. Those details aren’t available in the report.

What the financial claims show

Healthleap sells three-year contracts priced by licensed bed count and also uses outcome-based pricing. Meyer says the company contractually commits to delivering multiples of the contract price, using financial returns that hospital finance teams validate and attribute to its software.

He claims every customer has seen at least five times the contract price in “hard ROI,” with some seeing more than 20 times the price in annual total ROI.

At the Hospital of the University of Pennsylvania, Healthleap reports $23.8 million in annualized financial impact from its malnutrition program: $6.3 million in additional reimbursement and $17.5 million from shorter stays.

Those categories need separate scrutiny. Additional reimbursement is a revenue effect. Financial benefits attributed to shorter stays depend on assumptions about costs, available capacity, and how the hospital uses freed beds. Annualized impact doesn’t necessarily mean cash already realized over a completed year.

The figures are company-reported, and TechCrunch’s report doesn’t include the underlying attribution method or a controlled comparison. Changes in staffing, documentation, or clinical practice could affect both reimbursement and length of stay.

The business model helps explain the appeal to hospitals. Earlier identification can support care and documentation, while finance teams have identifiable outcomes to examine. Procurement should keep those measurements separate: improved reimbursement doesn’t by itself establish improved patient outcomes.

Expansion needs condition-specific evidence

Healthleap plans to use the funding for engineering, product, sales, and customer success. It ultimately wants to support more than 40 major conditions and extend into outpatient and home care.

Its programs for aspiration pneumonia, pressure ulcers, and congestive heart failure readmission risk are still undergoing further clinical validation, according to Meyer. Reusing extraction infrastructure can make new programs easier to build, but each condition brings different labels, time horizons, interventions, and costs of error.

Outpatient and home care introduce further complications. Records may arrive less consistently, and responsibility for acting on a flag may be less clear than on a hospital ward.

Healthleap has found a credible use for language models in hospital screening and expanded its customer base quickly. Evidence for each new program will need to keep pace. Hospitals evaluating the platform should ask to see the source behind a score, the workload created by its threshold, and the clinical results measured after someone acts on it.