Standard AI tools chop PDFs into isolated paragraphs and lose chronological state across encounters. CareNote connects every lab panel, prescription, discharge summary, and specialist consult into a persistent patient graph with a clickable PaperTrail linking every claim to its source page.
Active 3-Drug Prescribing Cascade & Chart Lore: Furosemide 40 mg + KCl 20 mEq ongoing for copy-forwarded "HFpEF" despite normal Echo/NT-proBNP (#p.2), triggered by Amlodipine 10 mg vasodilatory edema.
Unclosed Post-Discharge Hold (notable_absence): Temporary AKI hold on Empagliflozin 10 mg (#p.3) never closed despite renal recovery to baseline (Cr 1.12 mg/dL on 2025-03-04, #p.1).
Five weeks after Primary Care initiated Amlodipine 10 mg daily , Patient R.K. developed bilateral ankle edema that Urgent Care misattributed to new-onset heart failure (“HFpEF”) and treated with Furosemide 40 mg daily . Although confirmatory echocardiography and NT-proBNP (58 pg/mL, LVEF 62%) ruled out heart failure 6 days later, the unverified “HFpEF” label was copy-pasted forward across subsequent encounters .
Unnecessary loop diuresis for dihydropyridine CCB vasodilatory edema precipitated orthostatic syncope, prerenal AKI (eGFR 64 → 36 mL/min), and hypokalemia (K+ 3.1 mEq/L), prompting inpatient addition of KCl 20 mEq and a conditional hold on Empagliflozin 10 mg until Cr <1.2 mg/dL . Despite creatinine recovering to 1.12 mg/dL on 04 Mar 2025, Empagliflozin remains unresumed .
PaperTrail Inspector
Zero blind trust · Exact source anchor
Echo_Cardio_Nov2024.pdf
Diagnostic Report · 2024-11-25
Cross-Record Contradiction (Chart Lore vs. Echo):
“TTE & Biomarker Conclusion: LVEF 62%, normal LV diastolic filling (E/e' 7.2), normal LA volume, NT-proBNP 58 pg/mL (ref <125). No evidence of heart failure; ankle edema consistent with dihydropyridine CCB vasodilation.”
Standard AI tools slice clinical PDFs into isolated text chunks and retrieve by semantic similarity in an eternal present. Because vector search lacks a model of chronological validity intervals, it retrieves superseded regimens alongside active orders, confuses historical chart lore with current diagnoses, and cannot trace multi-encounter prescribing cascades.
Encounter records are broken into stateless fragments and indexed by keyword similarity. Retrieval repeats copy-pasted problem list labels (“chart lore”) and treats iatrogenic prescribing cascades as independent chronic conditions.
× Chunk 19 (Urgent Care): "Pedal edema: suspect HFpEF; start Furosemide 40mg"
× Chunk 44 (Discharge): "PMH: HFpEF. K+ 3.1; start KCl 20mEq. Hold Empagliflozin"
× Chunk 61 (Echo PDF): "LVEF 62%, NT-proBNP 58 pg/mL: no heart failure"
Aligns causal and epistemic edges across encounters and specialties. Links drug initiations to downstream symptoms, verifies copied diagnoses against primary diagnostic reports, and tracks conditional holds until closure.
Amlodipine 10mg → Edema → Furosemide → KCl → [PRESCRIBING CASCADE]
"HFpEF" Label vs. Normal Echo/BNP (#p.2) → [⚡ CHART LORE REFUTED]
Empagliflozin 10mg (Cr 1.12 < 1.2 threshold met) → [∅ UNCLOSED HOLD]
In medicine, a hallucinated dosage or misattributed date carries real danger. CareNote refuses black-box generation. Every asserted fact carries a direct audit anchor.
Multimodal AI reads the PDF visually, extracting values, units, reference intervals, and physical page numbers directly from the clinical record.
Facts accumulate into the patient graph, connecting changes across visits and detecting discrepancies between different specialists.
Every claim in the clinical synthesis links directly to its source PDF. Click the citation to inspect the original doctor's note and page.
Ask questions across accumulated clinical encounter records. Receive grounded answers with direct PaperTrail citations back to the exact source PDF page.
No. Echocardiography ruled out heart failure; Furosemide was initiated in a prescribing cascade:
Experience how persistent bi-temporal graph memory and clickable PaperTrail citations synthesize fragmented clinical encounter records.