Editing PHKG Business Opportunities (section)

== 4. GP Diagnostic Overview — AI Summarisation Layer on Kjernejournal ==

'''Concept:''' GP has a patient in front of them. The PHKG provides an AI-summarised, structured, ontology-backed view of the patient's history — not just a flat document list. AI highlights relevant history for the current complaint. GP diagnoses faster with fewer redundant tests.

'''Important positioning:''' This is NOT "first-ever cross-institutional view for Norwegian GPs." Norway already has '''Kjernejournal''' (national Summary Care Record, since 2017) — accessible from inside the GP's EHR, including overview of clinical documents from all hospitals (discharge summaries, x-ray results, lab results) via IHE-XDS protocol. '''Pasientens Legemiddelliste (PLL)''' is rolling out from 2024, providing national patient medication access. The product is an '''AI summarisation layer on top of Kjernejournal/PLL''', not a replacement for existing infrastructure.

=== How It Works ===

# Patient data flows through Kjernejournal (existing national infrastructure)
# Stamen's PHKG ingests Kjernejournal documents + unstructured clinical notes
# AI structures, codes (SNOMED CT), and summarises the full history
# GP sees a structured timeline + AI summary, not a flat document list
# AI highlights: relevant past conditions, current medications, trends, risk factors
# GP can ask natural language questions: "Has this patient ever had a cardiac workup?"

=== The Underlying Problem — With Sources ===

* '''History dominates diagnosis:''' Hampton et al. (1975) found history alone yields the correct diagnosis in 75-80% of cases. Replicated by Peterson et al. (1992, 76%) and Roshan & Rao (2000, 83%). If the GP has a better-structured history, consultation time shifts to treatment.<ref>Hampton et al. (1975): "Relative contributions of history-taking, physical examination, and laboratory investigation" — BMJ. https://www.bmj.com/content/2/5969/486</ref><ref>Peterson et al. (1992): https://pubmed.ncbi.nlm.nih.gov/1287004/</ref><ref>Roshan & Rao (2000): https://pubmed.ncbi.nlm.nih.gov/11273506/</ref>

* '''Time pressure is severe:''' Wilson et al. (2024, BMJ Open) found delivering recommended care would require >24 hours per GP per day.<ref>Wilson et al. (2024): BMJ Open — https://bmjopen.bmj.com/content/14/1/e077931</ref>

* '''Kjernejournal is flat:''' It provides a document store — discharge summaries, lab results, imaging reports — but not structured, ontology-backed, AI-summarised views. GPs still need to manually read and cross-reference multiple documents. The gap is intelligent summarisation, not access.

* '''Norwegian GP consultation time:''' '''No verified primary source found for a single average figure.'''

=== Clinical Safety — The Real Constraint ===

The evidence for AI chart summarisation is promising but has serious safety signals:

'''The promise:'''
* ChatGPT-4 adapted with in-context learning performed superior to physicians on both AI metrics and clinical expert evaluations.
* Best-in-class systems achieve 1.47% hallucination rate and 3.45% omission rate across 12,999 clinician-annotated sentences.

'''The risks:'''
* '''Omissions are worse than hallucinations:''' A 2026 study of an EHR-integrated AI chart review tool deployed nationally in the US found that missing/confusing information was more commonly reported than hallucinations. Errors of omission may be a larger threat than errors of commission.
* '''47% omission rate in ED summaries:''' A separate evaluation found 47% of 100 LLM-generated emergency department summaries omitted clinically relevant information.
* '''Automation bias:''' A 2025 study of Epic's GPT-4 summarisation tool at NYU Langone found 22.7% of providers reported sometimes skipping full-length notes in favor of the summary alone. If the summary omits a relevant finding, the tool causes the diagnostic miss it was supposed to prevent.
* '''Verification overhead erases time savings:''' PDSQI-9 validation required clinician evaluators averaging 10 minutes per evaluation. If GPs need to verify the summary against source notes, the time saving disappears.

'''Implication:''' The clinical safety case (omissions, automation bias, verification cost) is the binding constraint — not technology readiness or integration cost.

=== Competitive Landscape ===

'''Major competitors (wiki previously omitted):'''

* '''Epic's own AI chart summarisation (GPT-4 powered):''' Deployed nationally across US health systems including NYU Langone, scaled in 2024-25. This is the product Stamen would compete with directly. Source: NYU Langone study on automation bias.<ref>NYU Langone Epic GPT-4 summarisation study (2025) — automation bias in clinical AI chart review.</ref>

* '''Helseplattformen (Norwegian Epic deployment):''' Mid-Norway region. Adds complexity — potentially gives Epic a foothold for chart-summarisation features in Norway. Stamen would be competing against Epic's own AI features in Helseplattformen regions.

* '''Abridge, Nabla, Suki, DeepScribe:''' Primarily ambient documentation but expanding into chart review/summarisation. Well-funded.

* '''Microsoft/Nuance DAX Copilot:''' Same direction — ambient documentation expanding to chart summarisation.

* '''Tandem Health''' (Sweden/Norway) — AI clinical documentation with EHR access already negotiated. Nordic competitor with market presence.

* '''Cogstack''' (UK) — NHS clinical NLP, open source, extracts from unstructured notes. Research tool, not commercial product.

'''The framing "no existing GP tool does AI summarisation of full history" was true in 2023. By 2026 it is actively contested.'''

=== Norwegian GP EHR Landscape (corrected) ===

Norwegian GP EHRs are primarily '''CGM Allmenn''' (formerly WinMed/Profdoc), '''Infodoc Plenario''', and '''Pridok'''. DIPS is the dominant '''hospital''' EHR, not a GP EHR. Getting this wrong undermines credibility with Norwegian GP buyers.

=== Revenue Potential ===

* '''Per-practice subscription:''' €200-500/month per GP practice
* '''Norwegian GP market:''' ~5,000 GPs, ~2,500 practices
* '''Example:''' 100 practices × €300/month = €30,000/month = €360,000/year
* But: Epic's summarisation comes "free" with the EHR in Helseplattformen regions. Pricing pressure is real.

=== Challenges ===

* '''Clinical safety is the binding constraint:''' Not integration cost. Not technology readiness. A prospective clinical study showing the summarisation doesn't increase diagnostic misses is a prerequisite for GP trust and regulatory clearance.
* '''MDR / AI Act compliance:''' Any tool that highlights "relevant" history for the current complaint makes a clinical relevance judgment → clinical decision support under EU MDR → likely 12-18 months and significant capital for CE marking. Klinik has been navigating this for a decade; Stamen starts from scratch.
* '''Epic competition:''' In Helseplattformen regions, Epic's own GPT-4 summarisation comes bundled. Stamen needs to be demonstrably better, not just different.
* '''Integration:''' Connecting to CGM Allmenn, Infodoc Plenario, Pridok is technically complex. Each has different APIs, data models, authentication. 6-12 months per EHR.
* '''Automation bias:''' If GPs start relying on summaries without verifying, the tool becomes a liability. Product design must include verification nudges.

=== Verdict ===

'''The problem is real, but the opportunity is harder and narrower than it first appears.'''

* Norway already has Kjernejournal and PLL — the product is an AI summarisation layer on top, not first-ever cross-institutional access.
* The defensible niche may be narrower than "GP diagnostic overview": something like '''"structured ontology-backed summary for complex multimorbidity patients where omission risk is verifiable against the underlying PHKG"''' — leveraging the actual differentiator (PHKG structure) rather than competing on generic summarisation.
* '''Gating item is regulatory clearance and a published clinical safety evaluation''', not just data flow. Budget for a prospective clinical study, not just integration work.
* Epic's own summarisation tool, bundled with Helseplattformen, is the direct competitor. Stamen must be provably better on accuracy/omission rates, not just different on architecture.
* Year 2-3 timing is plausible but depends on completing clinical safety validation. Not before.