Unread Signal

Arkangel AI

Unread Signal

We turn clinical notes no health system reads into early warnings at population scale, across 305+ hospitals in 11 countries, before it's ever coded.

Bogotá, Colombia

Colombia

For-profit, including B-Corp or similar models

—

Every day, clinicians across Latin America and the Global South document what they observe: a cough that won't resolve, a family history mentioned in passing, social conditions shaping a patient's health. These free-text clinical narratives hold 80% of all clinical information — and no health system reads them at scale. Population health platforms, surveillance tools, and early warning systems run almost entirely on structured data: ICD codes, lab values, billing records. That's the 20% of the clinical picture that arrives late — often months after a patient's chart already indicated what was coming. In Latin America, where records are fragmented, coding is inconsistent, and primary care is overwhelmed, this gap is especially severe. The consequences are measurable. Patients with clear narrative trajectories toward uncontrolled diabetes, undiagnosed COPD, or progressive kidney disease are flagged only after hospitalization — not before. Globally, the WHO estimates that 80% of health data remains unstructured and underutilized. In the communities we serve — 305+ hospitals across 11 countries — the data to shift from reactive to anticipatory care already exists. It has always existed. It's written in the clinical notes that every system ignores.

Arkangel AI processes clinical narratives, the actual free-text notes clinicians write, and converts them into population-level health intelligence. We don't sample charts. We read all of them. We've built NLP models trained specifically on how clinicians in Latin America write: English, Spanish and Portuguese, full of abbreviations, institution-specific shorthand, and the compression that happens when clinicians are overwhelmed. Our published Pandora framework (IEEE Access, 2024) extracts diagnoses, risk factors, symptoms, social determinants, and temporal trajectories from real-world clinical text, catching what ICD codes miss. We then apply machine learning to stratify entire populations by disease risk, built on what each patient's clinician actually observed and documented. The result surfaces actionable signals: geographic clusters, demographic patterns, disease trajectory shifts, all before they appear in claims or coded diagnoses. In our pharma partnerships, we've identified patient cohorts 6 to 12 months before formal diagnosis. Currently operating at scale: 305+ hospitals, 11 countries across Latin America. Integrations with Epic, Azure, AWS, SAP. No data migration required, four to eight weeks to deploy. ISO 27001:2022 certified, HIPAA compliant. All processing within client infrastructure. Population sensing, not population surveillance.

Arkangel serves three interconnected populations. Patients in Latin America and the Global South (particularly those with chronic or underdiagnosed conditions like COPD, CKD, and uncontrolled diabetes) are the ultimate beneficiaries. These populations are disproportionately affected by late diagnosis in health systems where structured data is sparse and preventive care is under-resourced. By surfacing risk signals from the narratives clinicians already write, we enable earlier identification and intervention, before hospitalization, not after. Health system decision-makers: payers, hospital networks, and public health ministries, gain population-level intelligence they've never had access to. A payer managing 3.5 million members can identify undiagnosed disease cohorts from narrative signals; a ministry can detect district-level respiratory risk increases weeks before sentinel sites report them. Clinical teams are also direct beneficiaries: physicians in our partner institutions see their documentation treated as intelligence, not administrative burden. Their observations drive population health decisions rather than sitting unread in a database. The populations we serve (Spanish and Portuguese-speaking patients across fragmented health systems) are exactly those underserved by English-language, structured-data AI. Our NLP is built and validated on their clinical reality, not imported from contexts where it doesn't apply.