Machine learning to identify difficult to diagnose diseases |…

inTrigue Disease Learning fundamental to earlier disease diagnosis

Rare diseases take a long time to diagnose. Typically, they are diagnosed five to seven years too late.

This has changed with inTrigue because we learn new biomarkers from electronic health records that allow earlier intervention before coding (ICD9/ICD10/ICPC) has even occurred.

Earlier diagnosis means that prescriptions can be started sooner, leading to better health outcomes for patients and delivering a better return on investment for our clients.

Background to the inTrigue methodology

For rare and difficult to diagnose diseases

A range of challenges must be overcome in order to achieve earlier diagnosis and a full understanding of the disease. These challenges include the following.

Clinical coding in electronic health records

is seldom equivalent to the symptoms
is often incorrect
may, in fact, be completely absent

Symptoms may not be

fully understood
consistently documented

The disease is often understood differently in

countries
regions
disease understanding varies by type of clinician

This makes rare disease patient finding really hard.

Turning that challenge into an opportunity with inTrigue

The inTrigue approach is to re-learn.

Starting from

First principles
Originating data
Not derivative data

Challenging current assumptions to

Learn new things about the disease:
- Symptoms, comorbidities, natural history
Define new predictive Biomarkers:
- Cognitive, Digital, Physical
Create new much more effective Diagnostic Models for
- Earlier disease detection, more accurate disease detection, better outcomes

inTrigue Process overview

The first thing Volv does is see if the disease can be determined through any computational means at all, this is the predictability assessment
To diagnose diseases earlier we need to understand the nature of the disease in an agnostic way, taking into account all that we can discover about the condition. We call this disease modelling for prediction.
If we can know the disease very well, learnt through robust prediction, then we can identify cohorts better, meaning that we can know prevalence empirically. Volv can validate prevalence with a systematic methodology, enhancing our models further.
Once the patient cohorts have been identified, Volv uses the anonymised patient data to produce a bespoke patient journey explorer.
An outcome of the complete end to end process is novel biomarker discovery. This is new learning by Artificial Intelligence about the disease, embodied within the complex models produced.
However, while an artificial intelligence black box model can be very accurate indeed and can also be deployed in some settings, clinically interpretable models are of high-value, as they can be implemented in the clinical setting highlighted by the patient journey explorer to the best effect for patients.