Page 58 - AIH-1-2

Artificial Intelligence in Health AI in AD – Diagnosis and monitoring

imaging and AI for AD diagnosis, marking a substantial complex connections between clinical and serum measures
improvement in the field. The proposed method establishes in patients with mild-to-moderate AD.
a framework for automating the identification of skin Colonocytes or colonic epithelial cells have recently
conditions using MPT. garnered attention for their role in host-microbial
Activation-regulated chemokine (TARC/CCL17) and interactions. During gut dysbiosis, which is linked to a
immunoglobulin E (IgE) have served as biomarkers for number of chronic human disorders, colonocytes influence
AD in traditional approaches over the past few decades. 35-38 the composition and activity of the gut microbiota. The
45
Common techniques used in these investigations include diagnosis and prognosis of AD can now be achieved through
regression or correlation analyses between potential the integration and correlation analyses of gut microbiota
biomarkers and the intensity of AD symptoms, as well and host gene expression. 46,47 Notable correlations have been
as univariate research comparing AD patients to healthy observed, including those between IL-17 and Streptococcus
controls. However, accurate diagnosing and evaluating AD infection in AD, and between enzyme commission genes
solely based on a single biomarker are considered extremely and microbiota in inflammatory bowel illnesses. 48,49 Despite
challenging. Recent developments have ushered in the these advancements, few researchers have investigated ML
39
utilization of multivariate ML techniques to uncover prediction analysis based on the gut transcriptome and
50
hidden patterns between variables and develop more microbiota in AD. In recent work, Jiang et al. developed
reliable predictive models in a variety of studies, including an ML classifier for precise and automated AD detection
those in pain research. 40,41 The combination of multiple by utilizing the transcriptome of gut epithelial colonocytes
serum biomarkers, such as TARC, IL-22, and sIL-2R, has and gut microbiota data (Figure 3). With an average
improved the model’s capability to predict eczema area F1-score of 0.84, the classifier demonstrated accurate
and severity index (EASI) scores compared to relying on a discrimination and successfully predicted the risk of AD. It
single biomarker. In addition, the correlation coefficient was trained on data from 161 participants, including both
39
between the combined biomarkers and the disease AD patients and healthy controls. The research identified
severity surpasses that of the individual biomarkers. This three genes and three bacteria that are either directly or
method emphasizes the potential of combining multiple indirectly linked to AD, as well asa combination of 35
biomarkers for a thorough comprehension and prediction genes and 50 microbiome traits predictive for AD. These
of AD severity. results suggest that the discovered genes and microbiota
In a recent study, Lee et al. investigated the potential traits may provide fresh biological perspectives and serve
42
of employing a multivariate ML technique to develop a as useful biomarkers for early detection of AD. However,
diagnostic tool and severity prediction model for patients replication studies with different populations are necessary
with AD. The authors conducted phase I ML analysis, to validate these findings. The study represents a major step
wherein they collected multivariate data, divided it into toward the construction of an ML classifier for accurate
training and test sets, trained the models, estimated and automated AD diagnosis, utilizing gut microbiota and
prediction performance, and selected and estimated transcriptome data from gut epithelial colonocytes. The
features. Clinical and serological indicators 43,44 from a prior robust ML pipeline used in the study, which comprises
clinical study were combined. The results indicate that thorough procedures such as feature selection, model
the classification model significantly outperformed the selection, cross-validation, classification, and follow-up
random permutation model, boasting an area under the statistical assessments, enables accurate distinction based
curve of 0.85±0.10 and a balanced accuracy of 0.81±0.15, on omics data (Figure 4).
compared to 0.50±0.15 for the latter. Correlation analysis Dautović et al. developed an artificial neural network
51
unveiled a significant positive association between the (ANN) specifically designed for the automated diagnosis of
objective SCORing AD score (SCORAD) (r=0.53), AD, aiming to facilitate the diagnosis process. The network
measured and projected total SCORAD (r=0.43), and uses a feed-forward ANN with nine input parameters and
eczema area and severity index scores (r=0.58, each one output parameter for classification. After evaluating
p<0.001). Nevertheless, no discernible relationship was various configurations, the final design of the expert
observed between the measured and anticipated itch system chose a neural network with 15 neurons in a hidden
scores (r=0.21, p=0.18). The research encompassed the layer based on training results. Demonstrating impressive
creation and evaluation of multivariate prediction models, sensitivity at 95.62% and accuracy at 94.44%, the ANN
as well as the identification of critical characteristics using exhibits excellent performance in distinguishing AD from
a range of serum biomarkers. These results underscore the other skin disorders. However, it is imperative to recognize
potential of utilizing a multivariate ML approach to reveal that despite its high sensitivity, the comparatively lower

Volume 1 Issue 2 (2024) 52 doi: 10.36922/aih.2775