Pre-visit patient interview
Patient engagement →AI-led structured intake with the patient: history, ROS, social and environmental factors, red-flag triggers. Output structured for the consultation context.
Decision support. Each capability below is decision support for the clinician using ChironAI. The clinician evaluates, decides, and signs every output that enters the chart. ChironAI does not make a regulatory clearance claim; see Disclosures.
ChironAI CDSFull capabilities inventory
The capabilities Chiron reaches for, by workflow.
Roughly forty decision-support capabilities — not a menu you operate by hand, but the toolkit Chiron reaches into as it runs the encounter, calling each through a visible, named tool-call and drafting the result for attestation. They are grouped below by where they live in the consultation workflow. For specialty-depth treatment of any group, follow the specialty link in that section. The clinician remains the decision-maker on every output.
Pre-visit
Before the clinician walks in.
Risk-flag surfacing
High-priority flags surfaced from the patient interview before the clinician opens the chart. Red-flag patterns trigger AB 489-compliant clinical attention.
Patient consent capture
Explicit consent capture for AI-assisted workflows, per AB 3030 disclosure. Persisted with the visit record.
Reasoning
Clinical reasoning, structured and auditable.
Differential diagnosis support
Decision-support reasoning over presenting features with Bayesian confidence and qualitative tiers. Discriminating features called out per differential. The clinician evaluates and selects.
Clinical evidence synthesis
Live synthesis of canonical guidelines and peer-reviewed literature. Source-grounded with explicit guideline anchors.
Risk stratification
Wells (DVT/PE), GRACE (ACS), MELD, CHA₂DS₂-VASc, TIMI, HEART, others. The reasoning that justifies each score is shown.
Confidence calibration
Six-tier qualitative scale plus quantitative Bayesian percentages. “Cannot exclude” as a first-class state when data is insufficient.
Must-review-before-final gate
Compliance →Architectural AB 489 gate. Every AI artifact carries the non-dismissible review banner. Every PDF export carries the disclosure in the footer.
Interactive · Abductive chainStylised case · 4 steps
A 78-year-old male with palpitations, dizziness, and an irregular rhythm.
Walk the abductive chain the reasoner produces, one step at a time. Use Tab to focus controls, Enter to advance, and Esc to reset. The disclaimer above stays visible at every step.
- 01Observations gathered
- 02Hypotheses generated
- 03Red-flag scan
- 04Ranked differential
Step 01
Observations gathered
What the system sees, with provenance to its source.
The reasoner gathers structured observations before any inference is drawn. Each row carries its source so the clinician can audit provenance.
- Age / sex78-year-old male[intake]
- Chief complaintPalpitations + dizziness, intermittent, ~2 weeks[intake]
- Vitals (11:23)HR 122 irregularly irregular · BP 138/86 · SpO₂ 96% RA · Temp 37.0 °C[VS-1]
- ECG strip (11:28)Irregularly irregular rhythm, absent P-waves, narrow QRS[ECG-1]
- PMHHTN (10y), Type 2 DM (5y), prior CABG 2018[history]
- MedicationsMetformin 1g BID · Lisinopril 20mg QD · ASA 81mg QD[history]
- Recent contextNo fever, no chest pain, no syncope. One pre-syncopal episode standing from chair (yesterday).[intake]
Imaging
Radiology, structured and source-grounded.
Multi-pass radiology second-look
Radiology →Five-pass structured second-look review for radiology — structure, pathology, artifacts, missed zones, cross-window correlation — as a named ordered process. The radiologist drafts and signs the impression of record.
~35 frameworks, nine RADS systems
Around thirty-five named frameworks — BI-RADS, LI-RADS, PI-RADS, Lung-RADS, TI-RADS, ACR-TI-RADS, RECIST, PERCIST, ASPECTS, AO/OTA, and more, including nine RADS systems. Modality-appropriate framework selection.
Cognitive-bias counter-measures
The five-pass read carries explicit counter-measures against satisfaction of search, anchoring, and premature closure — built into the process, then resolved into a ten-section structured report.
Red-Alert discipline
Time-critical findings are architecturally separated from routine outputs. Notification path is distinct from the read.
Image-grounded reasoning
A dedicated vision model routes imaging through structured reasoning, with the reasoning streaming even as the study is read. Candidate findings cite the imaging series and slice they were observed on, surfaced for radiologist review.
Interactive · Vision routeThree modalities · No real imaging
The vision route, framework-aware on three modalities.
Switch between the three modalities to see the framework the reasoner applies, the schematic placement of candidate findings, and the structured impression the F5/reasoner drafts before the radiologist reviews and signs.
Schematic only. Abstract geometric shapes for illustration; not derived from any real imaging study.
Findings
- R breast: 14 mm spiculated mass at the upper-outer quadrant, new since prior. Suspicious morphology.
- L breast: stable oval density, benign-appearing on margin and density assessment.
- No suspicious calcifications either side.
Framework
Right BI-RADS 4 (suspicious abnormality) · Left BI-RADS 2 (benign).
Recommendation
Recommend image-guided biopsy of the right upper-outer mass. Continue routine annual screening of the left breast.
Draft impression — pending radiologist review and signature.
Radiologist reviews and signs. The reasoner drafts the structured impression; the radiologist of record edits, attests, and signs the impression that enters the chart.
Diagnostics
Labs, structured and pattern-aware.
Thirteen-pattern recognition library
Labs →Sepsis screen, AKI, thyroid panel, lipid panel, diabetic series, hepatic panel, renal panel, infectious panel, oncology panel, cardiovascular panel, and more — thirteen canonical patterns in all.
Ground-truth value extraction
Labs →Lab values are extracted by Azure Document Intelligence directly from the source report — a deterministic layer, separate from the reasoning engine. Chiron may label and interpret a value; it cannot invent a digit.
Twenty-two critical-value cutoffs
Every recognized pattern is cross-checked against twenty-two critical-value cutoffs before it reaches the ordering clinician.
Reference-range adjustment
Demographic adjustment for age, sex, pregnancy status, and applicable specialty context. The system does not flag normal pediatric values as adult abnormal.
Reflex testing recommendation
Where a pattern warrants additional testing, the recommendation surfaces with the guideline anchor that warrants it.
Documentation
Notes, structured and source-grounded.
SOAP source-grounding
Documentation →Every statement in the generated SOAP note traces to its underlying source field, source value, and source date. Visible to the clinician at review time.
Twelve-locale multi-language output
English, Spanish, French, German, Hindi, Mandarin, Arabic, Tagalog, Vietnamese, Korean, Portuguese, Russian. RTL support for Arabic.
Document versioning
Every signed document gets an immutable SHA-256 hash at signature time. Amendments are recorded as new versions; the original signed version stays verifiable.
Signature integrity
Document hash + signing clinician identity + timestamp + audit-chain entry. Any subsequent modification fails hash verification.
Bulk-sign workflow
Multiple documents reviewed and signed in one workflow. Each document still requires individual physician attestation; bulk-sign is a UI optimization, not a compliance shortcut.
Interactive · DocumentationBefore / after
The same visit. Two ways. Twenty-four minutes versus six.
Toggle between the raw, time-stamped notes a clinician types in the moment and the structured SOAP note the F5/reasoner produces — every clause source-grounded back to the intake field, the vital signs, or the ECG that warrants it. Stylised illustration, not a real patient.
With ChironAI™Clinician documentation time: ~6 minutes
Subjective
A 78-year-old male with hypertension, type 2 diabetes, and prior CABG (2018) presents with intermittent palpitations of two-week duration.
from intake question 4Reports a single pre-syncopal episode standing from a chair the day prior to evaluation.
from intake question 7Denies chest pain, dyspnoea, fevers, or recent immobilisation.
from intake ROS 2–6
Objective
Vital signs at 11:23: HR 122 irregularly irregular, BP 138/86, SpO₂ 96% on room air, temperature 37.0 °C.
from VS 11:2312-lead ECG at 11:28 demonstrates an irregularly irregular rhythm with absent P-waves and a narrow QRS complex.
from ECG 11:28Pulmonary examination clear bilaterally; no peripheral oedema; cardiac auscultation confirms irregular rhythm.
from exam fields
Assessment
New-onset atrial fibrillation with rapid ventricular response is the leading working diagnosis, consistent with the rhythm-strip morphology and the structural cardiac priors.
from differential rank 01Atrial flutter with variable conduction is held as a secondary consideration pending full 12-lead review.
from differential rank 02CHA₂DS₂-VASc score 3 (HTN, DM, age ≥ 75); HAS-BLED 1 — anticoagulation is indicated.
from risk-stratification panel
Plan
Rate control: IV metoprolol if blood pressure tolerates, with continuous telemetry monitoring.
from rate-control protocolAnticoagulation: apixaban once acute coronary syndrome is excluded; troponin pending.
from anticoagulation guidanceCardiology consult requested at the time of admission.
from consult request 11:34Disposition: admit to telemetry, NPO until cardiology evaluates, follow up in the morning.
from disposition field
Draft note, pending clinician review and signature. Every clause traces back to its source field. Document hash captured at signature time.
Prescribing
Medications, structured and interaction-aware.
Full medication schemas
Prescribing →Indication, mechanism, contraindication, drug-drug and drug-disease interaction, dose-adjustment guidance. RxNorm-anchored.
Four-tier drug interaction severity
Critical (contraindicated), Major (monitor closely), Moderate (caution), Minor (informational). With mechanism disclosure and recommendation per interaction.
Step therapy and prior auth flags
When a prescription would trigger a step-therapy requirement or prior auth under common payer formularies, the system surfaces the flag before the prescription is signed.
Pharmacology evidence anchoring
Each interaction cites the canonical pharmacology source that warrants the severity tier and recommendation.
Engagement
Patients, structured at their reading level.
Patient portal
Engagement →Patient-facing portal for pre-visit interview, post-visit follow-up, patient education, and consent capture. Mobile-first.
Six reading levels
Patient education output composed at kindergarten, elementary, middle-school, high-school, college, and professional reading levels. The system meets the patient where they read.
Multi-language patient education
Patient-facing content rendered in any of the twelve supported locales. RTL support preserved end to end.
Compliance and audit
The substrate that makes everything else defensible.
Append-only, tamper-evident audit trail
Security →HMAC + previous-hash audit log on every clinical action, written append-only. A database trigger blocks edits and deletes — even by the record’s owner — so the trail is tamper-evident by construction.
Ground-truth extraction
Lab values are read from the source document by a deterministic extraction layer, separate from the reasoning engine. The model interprets; it cannot fabricate the underlying fact.
AES-256 encryption + PHI de-identification
Sensitive fields are protected with AES-256 encryption, and PHI is de-identified in the reasoning pipeline, which works on the minimum it needs.
Deterministic clinical-alert engine
A deterministic engine of twenty-five rules fires clinical alerts on fixed criteria — predictable, testable, and independent of the model’s judgment.
Server-authoritative AB 3030 disclosure
Generative-AI authorship is disclosed on every AI artifact, enforced on the server. Non-dismissible by design — a client cannot dismiss its way past it.
ADMT notice and opt-out
AB 375 / CCPA / CPRA Automated Decision-Making Technology notice surfaces on first AI feature use. Opt-out controls in the user-preferences surface.
Architectural tenant isolation
Every database row carries a tenant boundary, and PostgreSQL Row-Level Security policies are defined at the database layer to isolate tenants independently of the application layer.
A note to the reader
See the full clinical workflow from intake to signed chart.
The capabilities above compose into a single end-to-end consultation flow. See how each step renders in the product.