Diagnosis of allergic diseases
The diagnosis or classification of allergic disease has been the area in
which AI has been applied most, an exemplary case of supervised
learning105–109. ML has used a wide range of data
sources to improve allergy or asthma diagnosis: text data from
electronic health records (EHRs), sound data of wheezes,image data from lung CT scans, or large-scale multi-omicsdata. The extraction of relevant clinical features from EHRs using NLP
has successfully diagnosed (childhood) asthma in discovery and
replication cohorts. In a study of a US birth cohort study, Seolet al. (2020) applied an AI algorithm to define asthma using
established predictive and diagnostic criteria in 8196 children. Of all
patients that met those criteria, 30% did not have a physician
diagnosis of asthma, signifying the potential for early disease
identification and population management with EHRs.
Additionally, several studies have investigated the potential of omics
data for diagnosis. One study developed an ML model that diagnosed
IgE-sensitized allergic disease in 16-year-old children based on nasal
cell DNA-methylation of only three CpG sites. External validation in an
independent cohort indicated the prospect of reproducible epigenetic
tests for diagnosis. Alag et al. (2019) pursued a similar
approach to diagnosing food allergy, where neural networks were trained
on blood epigenetic markers. The predictive markers were subsequently
associated with a 13-gene profile linked to immune response. This study
highlights the potential of novel diagnostic approaches to food allergy.
ML-based modelling of the component-resolved diagnostic multiplex array
data has shown that component-specific IgE responses to multiple
allergenic proteins are functionally coordinated and co-regulated, and
that the networks of interactions are associated with asthma diagnosis
and severity. Machine learning has also been used to predict disease
risk or persistence. In a prospective study of 704 children aged 2 to 13
months, unsupervised clustering on 16S rRNA data was used to identify
profiles of longitudinal changes in nasal airway microbiota that were
significantly associated with asthma risk at age seven. These results
affirm that the microbiome plays a vital role in the early development
of asthma and show promise for early identification and prevention
strategies. In another study, a supervised machine learning model was
able to predict asthma persistence in almost 10,000 patients diagnosed
before age 5 for persistence by age 10. The XGBoost algorithm delivered
the most robust performance (AUC=0.86), using clinically relevant
features such as the number of (non) asthma-related visits before age
five and noninvasive pulse oximetry data. The study was not
independently replicated, which is essential in pursuing clinical
support tools. Kothalawala et al. (2021) used data from birth
cohorts to train and validate two predictive models, CAPE and CAPP, to
predict the likelihood of asthma at school-age using predictors from 0-2
and 0-4 years of age, respectively. Predictive performance was
externally validated in the Manchester Asthma and Allergy Study (MAAS)
cohort. Support Vector Machine (SVM) algorithms provided the best
performance for both the CAPE (AUC=0.71) and CAPP (AUC=0.82) models, and
both demonstrated good generalisability in the replication cohort,
performing better than previous regression-based models.
AI guided image analysis has been performed to diagnose eczema. One
study developed a classifier of atopic dermatitis in multiphoton
tomography images, reaching over 97% accuracy through transfer
learning. Highlighted areas of interest in the images could support
clinicians in faster diagnosis.
Prediction of asthma exacerbations and hospitalizationsAsthma exacerbations are related to increased morbidity, mortality, and
healthcare use, yet these are challenging to predict. Several studies
have applied ML to predict exacerbations. In a large study involving EHR
data from 60.000 patients, researchers used different ML techniques in a
supervised setup to predict three exacerbation outcomes: oral
glucocorticoid bursts, ED visits, and hospitalization. The study
achieved a ROC AUC of 0.88 on the latter outcome, which is significantly
higher than the results of previous studies (AUC of 0.77); this was
replicated in an independent cohort. Important predictors for
hospitalization included oral glucocorticoid burst, inhaled
corticosteroid, and blood creatinine, the latter being unexpected.
Another study used self-reported daily home monitoring data of asthma
symptoms and peak expiratory flow, which were reduced in dimensionality
using PCA and then fed to various supervised ML methods. The best model
achieved a sensitivity of 90% and specificity of 83%, predicting
severe asthma exacerbations on the same day or up to three days in the
future. A more extensively validated example is Asthma-Guidance and
Prediction System (a-GPS), an AI tool to optimize asthma management.
A-GPS uses NLP on open text from EHRs to provide clinicians with the
most relevant clinical information. In a randomized control trial, the
tool significantly reduced the time for reviewing EHRs (11.3 to 3.5
min), but no significant change in clinical outcome (i.e.,
exacerbations) was observed. Sensor data from an electronic multi-dose
dry powder inhaler (eMDPI), such as inhalation volume and duration, has
also been utilized to predict exacerbations with a ROC AUC of 0.83.