In this early collaboration with Alnylam Pharmaceuticals and the PHARMO Institute for Drug Outcomes Research, Volv Global applied a novel machine learning methodology to a Dutch general practitioner database of 2.5 million patients – where no confirmed AHP diagnoses existed – to surface plausible undiagnosed cases. Expert clinical review confirmed 79% of the full candidate list as plausible AHP patients. The methodology’s disease-agnostic properties have since been validated across a growing portfolio of difficult-to-diagnose conditions.
Background
The disease
Acute hepatic porphyria (AHP) is a family of rare, genetic diseases characterised by potentially life-threatening acute attacks and, for many patients, chronic debilitating symptoms that significantly impair daily function and quality of life. The attacks can present with severe abdominal pain, neurological disturbance, and cardiovascular instability – a clinical picture that is easily confused with a wide range of more common conditions.
Despite frequent emergency department visits, GP consultations, and sometimes lengthy hospitalisations, patients with AHP are routinely misdiagnosed. Time to diagnosis has been reported at up to fifteen years – a prolonged journey that carries substantial clinical risk and a heavy personal burden.
The methodological challenge
Conventional patient-finding models in machine learning depend on a labelled dataset of confirmed cases. For AHP, and for many rare and difficult-to-diagnose diseases, that starting point does not exist at population scale. In a general practitioner database of 2.5 million patients, there were no records carrying clinical coding for AHP. Volv Global developed an approach designed to work from first principles, without that foundation.
Objective
- To identify potentially missed cases of AHP within a large, real-world general practitioner EHR database
- To demonstrate a replicable approach capable of operating without confirmed diagnostic labels
- To lay the groundwork for clinical deployment that could shorten the diagnostic journey for patients living with unrecognised AHP
Approach
Working with the PHARMO Institute for Drug Outcomes Research and a clinical porphyria expert at Erasmus Medical Center, Rotterdam, Volv Global developed a stepwise methodology to address the absence of confirmed AHP labels.
The PHARMO GP-EHR database contains de-identified records for 2.5 million patients arising from general practitioner encounters across the Netherlands, including demographics, ICPC diagnostic codes, ATC drug codes, and limited free-text clinical notes and laboratory results.
In the absence of any AHP-coded patients, Volv Global constructed a Silver Standard Labelled dataset of patients with clinical presentations similar to AHP – a maximum-similarity cohort drawn from across the full database. This candidate set was refined iteratively and assessed for predictability. A preliminary prediction model was trained from these clinically similar patients using supervised machine learning, then further optimised using unsupervised learning from unlabelled records, to mitigate the effects of unreliable labels.
To enable deployment within the data-secure environments typical of general practice, full models were abstracted into simpler versions capable of operating on modest on-site computational resources.
The resulting candidate list – the top 42 patients ranked by predicted probability of AHP, together with three controls – was reviewed in a blinded manner by an independent clinical porphyria expert practising in the Netherlands. Each candidate was scored as likely AHP, possible AHP, unlikely AHP, highly unlikely AHP, not AHP, or unable to assess.
All patient data were anonymised and provided to the PHARMO Institute by the data processor Stizon Foundation to ISO 27001 and NEN 7510 standards.
Results
Full list (top 42 candidates)
Of the 42 candidates, the expert assessed 17 as likely AHP and 16 as possible AHP – 33 plausible candidates in total, representing a precision of 79% (Precision@42).
Short list (top 20 candidates)
Of the top 20 candidates, 14 were rated likely AHP and four possible AHP, giving 18 plausible candidates – a precision of 90% (Precision@20).
Controls
None of the three control patients were rated as likely or possible AHP by the clinical reviewer.
These results were achieved by a model trained entirely without access to any confirmed AHP diagnoses.
What this demonstrated
For patients: AHP is typically diagnosed only after years of unexplained symptoms, repeated healthcare contacts, and avoidable harm. A model able to surface plausible undiagnosed cases from routine GP records represents a meaningful step towards shortening that journey – and towards ensuring that patients who are already in the healthcare system are recognised sooner.
For clinical teams: The approach produces a prioritised, clinician-reviewable shortlist, directly actionable within existing workflows without requiring changes to the data environment. The expert reviewer’s capacity to assess and validate candidates was built into the methodology from the outset.
For pharmaceutical partners: Understanding the true scale and distribution of an undiagnosed population is foundational to clinical development planning, patient registry design, and long-term programme strategy. This work showed how that understanding can be built even when confirmed cases at scale do not exist.
Wider significance
The most enduring finding from this project is methodological. The approach does not depend on confirmed diagnostic coding for the target disease. It learns from patients who share clinical features with the condition in question – a property that makes it directly applicable across the full spectrum of difficult-to-diagnose diseases.
This work was, to the authors’ knowledge, the first time a machine learning model for AHP had been built from clinically similar patients rather than confirmed cases. The same underlying principle now informs Volv Global’s proprietary methodology across disease areas including AATD, ARDS, Pompe disease, Fabry disease, HCM, and beyond.
Research details
This study was funded by Alnylam Pharmaceuticals. It is presented here as a record of Volv Global’s published research.
Authors: Rich Colbaugh, Kristin Glass, Christopher Rudolf (Volv Global SA); Ron Herings, Eline Houben (PHARMO Institute for Drug Outcomes Research); Janneke G. Langendonk (Porphyria Center, Erasmus Medical Center, Rotterdam, the Netherlands).
Links:
