2.0 PROJECT SUMMARY
Undertreatment of childhood asthma is prevalent and often the right treatment for incident cases is unknown
hence the widespread use of therapeutic trials as a treatment strategy. Two-thirds of incident childhood asthma
cases continue to have persistent symptoms even after treatment initiation. Missed opportunities for early
efficacious treatment contribute to increased risk of childhood asthma-associated morbidity (i.e., uncontrolled
asthma) that exerts a substantial burden on patients, families, and the healthcare system. However, clinical
decision-making tools needed to identify which child will benefit from which treatment at an early stage are
currently lacking.
This proposal is predicated on the notion that applying novel machine learning (ML) methodologies to
increasingly available electronic health record (EHR) risk/prognostic data can generate predictive analytics and
insights regarding childhood asthma treatment response. Clinicians can then use such insights toward effective
treatment decision-making at point of care, including more proactive and personalized treatment, for improved
patient-centered outcomes. Although risk and prognostic factors needed for treatment response prediction are
often embedded in EHR, this information is sometimes overlooked by clinicians. In busy pediatric clinics, active
EHR review to identify such factors to inform treatment decisions can be costly, time consuming, error-prone,
and infeasible.
To address these challenges and technological gap, we propose to develop, validate, and evaluate a childhood
asthma Passive Digital Marker for treatment response prediction (PDM-TR), that is, a ML algorithm that can
retrieve and synthesize pre-existing `passively' collected mother-child dyad risk/prognostic data in `digital' EHR
to provide an objective and quantifiable `marker' of treatment response.
We hypothesize that when applied to risk/prognostic EHR data derived from incident asthma cases exposed to
first-line treatments, our PDM-TR will predict asthma control at 2-3 months with high accuracy (=80 sensitivity
and =80 specificity). The PDM-TR will `learn from existing EHR data' to predict whether a specific treatment may
be successful (i.e., achieve asthma control) for a given individual with a specific set of attributes (i.e., asthma
risk and prognostic factors [e.g., history of allergy sensitization, eczema, demographics, lung function, body mass
index]). Applying our novel PDM-TR in-real time to readily available EHR data could contribute towards the
development of a timely, accurate and scalable approach to inform personalized childhood asthma treatment at
point of care.