Page 60 - AIH-1-2

Artificial Intelligence in Health AI in AD – Diagnosis and monitoring

Table 1. Artificial neural network system performance employed an ANN to examine the association between
the severity of AD and exposure to air pollutants and
Predicted Predicted Output class environmental factors. Their results revealed a robust
positive output negative
output association, with the severity of AD lesions increasing by a
Actual True positive False negative 1 – Subjects with considerable 200% in response to an increase in the diurnal
positive: (TP):153 (FN): 7 disease temperature range, defined as the difference between the
160 Output class highest and lowest temperatures of the day. By predicting
0 – Healthy subjects disease severity based on environmental parameters,
Actual False positive True negative Accuracy: 94.44% the ANN exhibited promise in providing patients with
negative: 20 (FP):3 (TN): 17 early warnings to avoid potential irritants, in line with
Σ = 180 Sensitivity: Specificity: 85% F1 score: 0.9684 the overarching objective of predictive model-informed,
95.62% tailored health-care actions.
MCC: 0.7454
Abbreviation: MCC: Matthews correlation coefficient. Neural network algorithms provide a reliable and
non-invasive method for classifying AD, frequently
demonstrating efficacy in assessing lesion severity.
representative datasets, ensuring the interpretability This capability eliminates the need for direct clinical
and explainability of AI models, addressing ethical or specialized dermatologist intervention by allowing
considerations, and conducting rigorous clinical validation. individuals to remotely monitor their condition using the
The fulfillment of these requirements is essential to ensure cameras on their mobile phones. However, it is imperative
the reliability and generalizability of proposed diagnostic to acknowledge that ANNs are not designed for constant
tools for AD. observation, and existing models generally attain a
maximum documented accuracy of 90%. Therefore,
5. AI for monitoring AD future research should prioritize enhancing accuracy and
Clinical professionals’ subjective visual inspections are developing user-friendly mobile devices or applications to
frequently used to determine the severity of AD, which enable patients to confidently assess their condition using
introduces significant inter- and intra-observer variability, cutting-edge algorithms. In addition, additional scientific
especially in varied clinical study settings. In an attempt evidence is required to determine how the use of emollients
to standardize and automate the diagnosis of AD severity, or moisturizing creams affects the sensitivity of ANN.
52
Pan et al. presented EczemaNet, a CNN computer vision These studies could yield important information regarding
pipeline. EczemaNet operates by initially identifying areas whether these neural network systems can be used to
affected by AD in images, and subsequently generating measure the effectiveness of dermatitis treatment progress.
probabilistic predictions regarding the severity of the Table 2 provides an overview of key information from the
condition. To generate its final predictions, EczemaNet included studies, while Figure 5 illustrates the various data
uses ensemble approaches, including crops, ordinal formats used in the previously described studies.
categorization, and transfer and multitask learning.
During evaluation in a published clinical trial, EczemaNet 6. Implications of AI in evaluating pruritus
exhibited minimal root mean square error and well- in AD
calibrated prediction intervals. The research demonstrated The sensation of itching, medically referred to as pruritus,
the effectiveness of CNNs in treating non-neoplastic is a complex problem that significantly affects one’s overall
skin conditions, especially when dealing with medium- quality of life. The persistent and intense prickly discomfort
sized datasets. This finding highlights their potential for associated with AD can cause mental health issues. These
delivering an objective and more effective assessment of problems may manifest as increased activity levels,
AD severity, which is a development with greater clinical generalized anxiety, and, in certain cases, major depressive
significance compared to simple classification techniques. disorders. 55-57 Despite the profound detrimental effects of
Padilla et al. used the MobileNet architecture, a CNN, pruritus in AD, the lack of standardized and established
53
to successfully distinguish between psoriasis and AD. techniques for objectively evaluating it poses a challenge
They utilized publicly accessible dermatology to train the to physicians and researchers. Pruritus is inherently
58
network. In a real-world experiment involving a Raspberry subjective, with diagnosis primarily reliant on patient
Pi camera and 30 subjects, the model successfully classified reporting. To address this issue, numerous metrics and
psoriasis with an impressive 90% accuracy rate and AD surveys have been developed. One such commonly used
with an 88% accuracy rate. In a related study, Patella et al. 54 tool is the Peak Pruritus Numerical Rating Scale (NRS),

Volume 1 Issue 2 (2024) 54 doi: 10.36922/aih.2775