HealthLeap

The model is the product. HealthLeap Nutrition reads every adult inpatient's chart each day, scores each patient's nutrition risk and writes the score into the EHR flowsheet, sorted so dietitians see the highest risk patients first. The company's peer reviewed study describes the model as gradient boosted decision trees that use a large language model to extract features from clinician notes.

The company wraps it in implementation, change management and outcomes services, but those exist to deliver the model's output and nothing is sold without it. Malnutrition is the first condition; delirium is described as built and tested, and aspiration pneumonia, heart failure readmission and pressure injury models as planned.

A dedicated clinical use and safety page (updated October 2025) states what the tool may and may not do. It is clinical decision support for registered dietitians, for adult inpatients only, not for pediatric, outpatient or emergency decisions, and not for time critical decisions. It is not a diagnostic tool, gives no treatment recommendations, must not be the sole basis for determining nutritional risk or malnutrition, and supports rather than directs decisions; each output shows the contributing data so a dietitian can check it, local policies prevail, and a reporting channel for safety concerns is given. Nothing is acted on automatically: the score reorders the dietitian's list.

The threshold that places a patient on the priority list is not published, nor is what a site sees about the patients the model ranks low; the study found about half of malnourished patients were not flagged on the first day.

The approach and a supplier are named. The company's study describes the model as gradient boosted decision trees, with features drawn from demographics, weights, laboratory values, diagnoses, problem lists, medications, billing codes and prior admissions, plus features extracted from the longest clinician note around admission by OpenChat-3.5-0106, an open source fine tune of Mistral 7B, prompted with questions built on the ASPEN and Academy criteria and the GLIM framework. Training data is characterized: 77,855 patients at Cedars-Sinai, with later admissions held out for validation and testing, and albumin and BMI as the strongest features. Each output shows its contributing factors.

No version or update practice is published, and the product description does not match the study: the clinical use page describes structured data and evidence based risk criteria with rule logic, and mentions no language model or notes.

The model layer is named. The company's study says it used OpenChat-3.5-0106, an openly released fine tune of Mistral 7B, to read clinician notes, alongside its own gradient boosted model, which points to a model the company runs itself rather than a third party service receiving notes.

Nothing else in the chain is named: no hosting provider, region, subprocessor list, or statement that the same language model runs in the product today, and the website privacy policy's vendors and analytics partners cover the website, not the platform.

Peer reviewed and single site. Bernstein and colleagues (Applied Clinical Informatics, 2025, PMC12618146), led from the company, studied 166,841 admissions (106,449 patients) at Cedars-Sinai from January 2019 to September 2022, retrospectively, with a later out of time test set of 14,094 admissions. AUROC was 0.92 on the first day and 0.95 using each patient's highest score during the stay, against discharge coded malnutrition, with similar results against dietitian records. At a threshold matched to the nurse screening tool's flag volume, the model was more sensitive on the first day (0.52 against 0.35 with dietitian records as reference, 0.49 against 0.24 with coding), and patients it flagged had higher readmission and mortality. It flagged malnutrition before the dietitian documented it 72.7 percent of the time and after it 10.4 percent, and missed 16.9 percent; the advance was a median of 1.3 days and a mean of 4.0. The company's home page quotes the 4 day mean.

The evidence is retrospective at one hospital, which the authors name as a limitation, and the impact the company markets (an annualized $23.8 million and 8,632 bed days at the Hospital of the University of Pennsylvania, $11 million and 39 percent more diagnoses at Cedars-Sinai) comes from case studies whose method sits behind a form. Named partners are Cedars-Sinai, Emory, Houston Methodist, Intermountain, Northeast Georgia Health System, Penn Medicine and UMass Memorial.

What reaches the model is described; what happens to it afterward is not. The company's privacy policy says in terms that it does not cover the AI platform, and its website terms defer to each customer's business associate agreement and data processing agreement, so nothing published states retention, training use or de-identification for patient data.

The validation study describes the data path: structured record data (demographics, weights, laboratory values, diagnoses, problem lists, medications and billing codes) feed the model, and the longest clinician note from the day before admission to the first midnight is read by an open source language model (OpenChat-3.5-0106, a fine tuned Mistral 7B) to extract features; the paper says generating a score takes under an hour including data transfer, which implies data leaves the hospital's system. Every adult inpatient chart is read daily, so the company processes records for patients it never flags, and free text notes hold material unrelated to nutrition.

Business associate status is stated in passing and nothing published supports it for the service. The website terms say that customers with a master services agreement, business associate agreement or data processing agreement with the company are governed by those documents for the enterprise platform, and that protected health information may be sent only by institutions with an executed business associate agreement.

But the only privacy policy published says in terms that it does not cover the AI platform, and no agreement, template or summary of its terms is published. Two scope points matter: the platform reads every adult inpatient chart daily, including patients it never flags, and its output feeds documentation and coding downstream.

Nothing is published: no SOC 2 report, ISO 27001 certificate, HITRUST certification, penetration test, trust center or description of controls for the platform. The website privacy policy's reasonable efforts language covers the website, and the clinical use page places data security and access controls on the customer site. The access at stake is broad: the platform reads every adult inpatient chart daily, including free text notes, and writes into the record.

The position is stated plainly. The clinical use page says HealthLeap Nutrition is not an FDA cleared or approved device, and that it is designed to meet the exclusion for non device clinical decision support under section 520(o)(1)(E) of the FD&C Act: it uses only existing EHR data, does not acquire or process medical images or device signals, supports rather than directs decisions, and shows the data elements behind each output so a clinician can review the basis independently. The scope matches: registered dietitians, adult inpatients, not pediatric, outpatient, emergency or time critical use.

One point for a buyer to check. That page describes the product as using structured data and evidence based risk criteria with rule logic, while the company's published model scores patients with gradient boosted trees and reads clinician notes with a language model, and its home page shows a numeric risk score ranking patients. Whether each output's basis can be reviewed the way the exclusion requires depends on which of those runs on your site.

Subgroup results are published, with the weak spot visible. The company's validation study (Bernstein and colleagues, Applied Clinical Informatics, 2025; the corresponding author uses the company's address, four authors are employees and three are advisors, all declared) reports performance separately for intensive care and other inpatients at Cedars-Sinai. On the first day of admission, against discharge coded malnutrition, sensitivity was 0.28 in the ICU against 0.50 elsewhere, and AUROC 0.81 against 0.92; against dietitian recorded malnutrition, ICU sensitivity was 0.26 against 0.53. The paper also examined calibration by race and sex, inspecting predicted against actual cases within risk deciles, and reports no substantial over or under prediction for any group; those results are in supplementary figures. The company cites the study on its home page as the validation of its product.

The study is one hospital, so how the model behaves across populations elsewhere is untested in public. The race and sex result is a visual calibration check rather than published figures. And the company's clinical use page describes the product as using structured data and evidence based risk criteria with rule logic, while the study's model reads clinician notes with a language model, and nothing published says which of the two runs at a given site.

A published error rate the company stands behind. Its home page cites its own study (Bernstein and colleagues, Applied Clinical Informatics, 2025), which reports, on a held out test set of 14,094 admissions at Cedars-Sinai from 2019 to 2022, with the risk threshold set so the model flagged as many patients as the nurse screening tool it replaced: first day sensitivity 0.49, specificity 0.97, positive predictive value 0.55 and negative predictive value 0.96 against discharge coded malnutrition, and 0.52, 0.96, 0.37 and 0.98 against dietitian recorded malnutrition. So on the first day about half of malnourished patients are flagged and, against dietitian records, about two in three flags are not confirmed. The company also says it offers a contractual return on investment floor and puts its own money at risk on the outcome, a commitment tied to reimbursement results rather than to the accuracy of the output; its terms are not published.

The company's clinical use page says results may be incomplete, delayed or inaccurate, that performance figures are specific to the study and will vary by institution, and that sites should evaluate the tool locally.

Integration is described by effect and nothing is named. The company says it reads every adult inpatient chart daily and writes a nutrition risk score into the EHR flowsheet each morning, so dietitians see a sorted list without a new application or login, and that no extra documentation is needed. No record system, interface standard, marketplace listing or integration document is named on its pages, even though its seven named partners are large health systems.

The company names seven health system partners (Cedars-Sinai, Emory, Houston Methodist, Intermountain, Northeast Georgia Health System, Penn Medicine and UMass Memorial), and a first quarter case study at the Hospital of the University of Pennsylvania (updated August 2026). Where the platform runs is not stated.

The validation study says generating a score takes under an hour including data transfer, which implies record data moves to the company's environment, and the website's privacy policy says the website is hosted in the United States, but no hosting model, region or residency commitment is published for the platform, and no on premises option is described.

No price is published, and the posture can be established before contact. The company sells the model with implementation, change management and outcomes accountability services, offers finance buyers a contractual return on investment floor, and says it puts its own money at risk on the outcome.

Its value case is published in figures: roughly $5 million in incremental reimbursement for every 1,000 additional reimbursable malnutrition cases, and case study results of $23.8 million annualized at one Penn Medicine hospital and $11 million a year at Cedars-Sinai. What is missing is the price itself and the unit of charge.

The case rests on reimbursement from documented malnutrition, and nothing says whether fees are tied to captured revenue; a share of recovered reimbursement would give the company a stake in how many diagnoses the tool helps document.

Coverage is named with its limits: adult inpatients (18 and over), used by registered dietitians, and not for pediatric, outpatient, emergency or time critical decisions, per the company's clinical use page. The live condition is malnutrition; a delirium and toxic metabolic encephalopathy model is described as built and tested, and aspiration pneumonia, heart failure readmission and pressure injury as next. Validation sits behind part of this.

The published study covers one hospital and shows markedly weaker first day sensitivity in intensive care (0.28 against 0.50 elsewhere), which the product's scope does not carve out, and no validation is published for the other conditions. The company's own page tells sites to evaluate the tool locally before relying on it.