FeelBetter

Two of the platform's four tools are the model and the other two deliver its output. Insight stratifies an entire population by risk of near term medication related deterioration, and Action generates the specific regimen changes a pharmacist should consider, while Navigator and Impact handle workflow and post intervention monitoring.

Remove the learned components and what remains is a pharmacist worklist with no basis for deciding which of a hundred thousand patients to open first, which is the entire proposition. The peer reviewed evaluation at Brigham and Women's Hospital describes it as a machine learning system and assessed the model's output rather than the workflow around it. Pharmaco-Clinical Intelligence is a term the company coined for its market position, not a description of a method.

The oversight chain is sound and conventional. Recommendations go to a clinical pharmacist who reviews them, and any medication change still requires a prescriber to act, so two licensed humans stand between the model and the patient. The weak point is a second order effect of the product's economics, and it applies to polypharmacy review products generally, not only to this vendor.

Part of what is being sold is throughput, with the company's own site citing a reduction of roughly 30 percent in average review time per complex patient and its materials emphasizing expanded clinical pharmacist capacity. Human review that becomes faster and covers more patients is human review that is thinner per case, so the oversight that justifies the product's autonomy is the same resource the return on investment depends on consuming.

A peer reviewed evaluation, which FeelBetter cites, names the approach: a machine learning system that uses medications, demographics, diagnoses, laboratory results, health care utilization patterns and costs to place each patient in one of 22 risk levels by risk percentile for emergency department visits or hospitalization, and generates medication recommendations for patients at medium and high risk (Seger, Rotenstein et al., American Journal of Managed Care, 2024;30(8):e233-e239).

FeelBetter's own site describes AI models that synthesize and sequence each patient's pharmacy and clinical data, and markets its approach under a coined term, Pharmaco-Clinical Intelligence, that names a market position rather than a method. No model family, version or update practice is given. Nothing describes how the model is retrained, how a customer learns that it changed, or what risk threshold a deployment uses to decide who reaches a pharmacist.

The architecture is described: FeelBetter says its AI models synthesize and sequence each patient's pharmacy and clinical data, and the peer reviewed evaluation in the American Journal of Managed Care (2024;30(8):e233-e239) states that the system uses medications, demographics, diagnoses, laboratory results, utilization patterns and costs to assign each patient one of 22 risk levels.

No model provider, hosting provider or subprocessor is named, and no retention, de identification or minimum necessary terms are stated. The footprint is population wide by design: a deployment means the vendor holds a longitudinal medication, clinical and claims picture for patients who have had no encounter with the product, and a risk model of this kind improves as its training population grows, so whether customer data informs models used for other customers matters, and nothing published answers it.

FeelBetter's privacy policy notes that service providers may be located in other countries, and the company operates from Boston and Tel Aviv, yet nothing in writing says where United States patient data is processed.

The peer reviewed work is real, and the distinction between what it measured and what the company claims is the important part. Investigators at Brigham and Women's Hospital published a retrospective cohort of 108,817 older patients with multimorbidity in the American Journal of Managed Care (Seger, Rotenstein et al., 2024;30(8):e233-e239), reporting that the system stratified emergency department and hospitalization risk and that its medication warnings were correct in 89.2 percent of cases and helpful to a clinician in 97.3 percent, judged by two pharmacists and a physician.

Those are judgments about the appropriateness of recommendations, not measurements of what happened to patients after acting on them. The study's principal investigator has said a prospective study with FeelBetter is under way at Brigham (Healthcare Innovation, September 2024). Operational results are vendor reported, including a pilot across eight Americare long term care facilities with 370 residents. The published research stops short of an outcome study; the prospective results are what would close that gap.

The platform synthesizes longitudinal pharmacy records, clinical data, utilization and cost across a covered population rather than examining one prescription at the moment it is written, and the published evaluation alone covered 108,817 patients. FeelBetter's privacy policy covers its website and offers only general assurances of safeguards and retention no longer than necessary.

No retention schedule, de identification practice, minimum necessary scoping or statement of whether customer data informs model development across institutions is published, and that last question matters most for a risk model whose value grows with the population it has learned from. The company operates from Boston and Tel Aviv, and nothing states where United States patient data is processed. No safety engineering is published: nothing describes guardrails on the recommendations or how a safety event is handled.

FeelBetter publishes no HIPAA statement, business associate position, agreement terms or compliance page. Its privacy policy covers its website, is governed by Israeli law, and describes website data only; it does not mention protected health information. No document that reaches the product states a HIPAA status, and no parent company's policies fill the gap.

The deployments at Atlantic Health System, Brigham and Women's and Americare will each rest on a business associate agreement, and that remains an inference until FeelBetter publishes its position. Nothing published says how an Israeli headquartered company processes United States protected health information.

No SOC 2 attestation, HITRUST certification, ISO 27001 certificate, trust center, penetration testing cadence or vulnerability disclosure policy is published. The privacy policy's data security section says FeelBetter maintains physical, technical and administrative safeguards and tests them periodically, which is an assertion with nothing independent behind it. The deployments suggest that customer security reviews were passed, but those reviews are private. For a platform holding longitudinal medication and clinical histories for entire senior populations, a published attestation is the single biggest thing missing.

FeelBetter claims no clearance and publishes none, which is consistent with a product that stratifies risk and proposes options to a clinical pharmacist rather than computing a dose or acting on a patient. That position is available under the statutory exclusion for non device clinical decision support, and the pharmacist plus prescriber review chain is the kind of arrangement the exclusion contemplates.

The company states no regulatory position anywhere, which matters because the boundary is contested. A model recommending discontinuation of a medication, which is what deprescribing means in practice, is making a therapeutic proposal rather than surfacing a reference fact, and nothing published says how the vendor characterizes that distinction.

Nothing published describes how the model's behavior is governed or tested across the patients it ranks: no subgroup performance, bias assessment, governance documentation or post deployment monitoring statement, and FeelBetter's site carries no responsible AI framework. The distributional question is concrete here because the published evaluation states the model's inputs include health care utilization patterns and costs.

Utilization and cost reflect access to care as well as illness, and risk models built on them can under identify need in groups that historically received less care. For a model that decides which older patients a pharmacist reviews first, no ranking is reported by race and ethnicity, sex, age band, dual eligibility or payer. The evaluation cohort of 108,817 patients is large enough to answer that, and nothing published says whether it was examined.

The mechanism that lets someone catch a wrong output is the review chain: recommendations go to a clinical pharmacist, and any medication change needs a prescriber to act. The peer reviewed evaluation in the American Journal of Managed Care (2024) tested exactly that output, with two pharmacists and a physician independently judging the system's warnings and recommendations: warnings were correct in 89.2 percent of cases and useful to a clinician in 97.3 percent, which amounts to a traceable error rate for the recommendation engine.

Nothing stands behind the output: no warranty, indemnity or remediation commitment, and no published risk threshold. The output is a prioritization of which patients a pharmacist reviews, so the error that matters most is the patient never surfaced, and that patient generates no record.

Integration is asserted at a general level and confirmed only by inference. The platform is described as fitting into existing claims and electronic health record workflows, and three live deployments across a health system, a long term care operator and an Israeli health maintenance organization establish that data flows in practice.

What is missing is any specific evidence a buyer could check: no named electronic health record certification, no Epic or Oracle marketplace listing, no FHIR conformance statement, no published API and no description of whether ingestion is claims based, clinical feed based or both.

Since the product depends on assembling a longitudinal pharmacy and clinical picture across sources, the ingestion architecture is a material procurement question and it is currently unanswerable from public material.

The product is delivered as a hosted platform and operates for customers in both the United States and Israel, which is as much as public material establishes. No hosting provider, cloud region, data residency commitment or on premise option is named, and nothing states where United States patient data is processed relative to the company's Israeli engineering base. Third party company profiles list infrastructure details that the vendor itself does not confirm.

For customers in long term care and accountable care organizations, both of which carry their own contractual data obligations upstream, this is a gap likely to surface during security review rather than before it.

Pricing is undisclosed, and the value side is only partly presented too. The company's own site lists its impact metrics as labeled categories, savings per high risk patient per month and average review time per complex patient, with one carrying a figure and the other left as a heading, so even the return side is only partly quantified while the cost side is entirely absent.

No price, unit, pricing basis, contract shape or implementation fee guidance is published, and nothing indicates whether the commercial unit is the covered life, the high risk patient, the pharmacist seat or the intervention. For a product sold into accountable care organizations on a total cost of care argument, the pricing unit determines whether the economics work at all.

Coverage is narrow by population and unusually broad by buyer. The population is consistent throughout, older adults with multiple chronic conditions and complex regimens, which is the group where polypharmacy causes the most harm and where the intervention has the clearest value.

Against that focus the platform has been deployed across three structurally different settings: an accountable care organization model at Atlantic Health System, long term care at Americare, and a health maintenance organization at Leumit Health Services in Israel, each with different data, different clinical staffing and different incentives. That breadth of buyer is harder than breadth of specialty and it is evidence the platform is not built around one customer's data model. The specialty scope, though, is essentially internal medicine and geriatrics, and the named footprint remains small.