Page 57 - AIH-1-2

Artificial Intelligence in Health AI in AD – Diagnosis and monitoring

text, while NLU is committed to understanding textual enhancement of categorization accuracy is feasible by
material. NLG encapsulates recent advances in large concentrating on wave number bands with a variable
language models, exemplified by OpenAI’s freely available importance in projection (VIP) score of ≥1. In addition to
Chat Generative Pre-trained Transformer. These bolstering the model’s accuracy, the VIP score facilitates
27
advancements highlight the evolving landscape of NLP the identification of important Raman spectroscopic
and its important role in the advancement of language- signatures associated with proteins, lipids, and nucleic
related applications. acids, which can serve as biomarkers for therapeutic and
In recent years, the adoption of multimodal techniques clinical evaluation of AD patients’ skin health. Using CRM
in algorithms has surged, driven by the utilization of diverse and multivariate analysis, this quantitative method of
data sources for training. Given the inherently multifaceted assessing skin inflammatory disorders such as AD offers a
nature of medicine, where doctors must interpret a wide viable path for next-generation diagnosis, departing from
range of data, including genetic information, laboratory the subjective scoring systems currently used in clinical
results, clinical notes, and radiological images, these practice. The presented study describes a novel diagnostic
multimodal approaches have gained prominence. method specific for AD using CRM and multivariate
The latest strides in this discipline focus on building analysis. This non-invasive method will provide a
more reliable models and algorithms by leveraging the new approach for molecular-based evaluation of skin
abundance of readily available data. Noteworthy examples conditions. Nevertheless, several challenges need to be
28
of these multimodal technologies include Med-Flamingo, addressed, such as sample size and diversity, independent
30
29
LLaVa-Med, Med-PaLM Multimodal (Med-PaLM M), dataset validation, clinical utility assessment, CRM
and MiniGPT-4. At the core of these technologies, they standardization across different laboratories, patient data
31
lie foundation models (FMs), which undergo training privacy and informed consent ethics issues, equipment
on a variety of unlabeled datasets before being adjusted accessibility, and cost. Regulatory approval for CRM
for certain downstream applications. One particularly technology’s widespread use is also necessary. Overcoming
32
intriguing aspect is the ability of these models to absorb these issues will improve the power and generalizability of
vast amounts of information from large datasets and this innovative diagnostic protocol for AD. Furthermore,
subsequently apply this knowledge to specific applications, the application assures the availability of significant
including those within the medical domain. This pattern datasets and ensures the repeatability and reliability of the
represents a dynamic movement in the direction of using model.
multimodal techniques to improve performance in medical Multiphoton tomography (MPT) has previously
applications. demonstrated its utility as a diagnostic tool in dermatology.
However, MPT data analysis has remained time-consuming
4. AI for the diagnosis of AD and operator-dependent. In a study conducted by
34
Accurate dermatological diagnosis and treatment of Guimarães et al., the potential of using AI for diagnosing
AD hinge on the quantitative evaluation of the disease, AD from MPT images was substantiated. AD system was
which emphasizes the molecular composition of the skin developed to discern images containing living cells and
using non-invasive techniques. Confocal Raman micro- performs subsequent diagnostics accurately and reliably,
spectroscopy (CRM) serves as a tool for assessing the skin’s thus eliminating the need for human operators. The study
biomolecular composition. Nevertheless, deciphering has demonstrated the potential of completely harnessing
complex Raman spectroscopic signals requires multivariate MPT through a CNN-based, fully automatic method.
analysis. Dev et al. have presented a novel approach to CNNs were trained and fine-tuned using 3663 MPT images,
33
classifying AD from healthy individuals by combining including morphological and metabolic information from
CRM with multivariate analysis, more precisely, partial both AD patients and healthy individuals. The primary
least squares discriminant analysis (PLS-DA). While the objectives were to identify live cells and diagnose AD,
current PLS-DA classification model is designed for binary irrespective of the imaging layer or location. Impressively,
classification, there is potential to explore its applicability the suggested algorithm successfully diagnosed AD in
for multiclass categorization based on the severity of 97.0±0.2% of the images containing live cells, with a
eczema illness. The ML-aided PLS-DA classification sensitivity of 0.966±0.003, specificity of 0.977±0.003,
approach used in the study simplifies dimensional and F-score of 0.964±0.002. The interpretability of the
reduction, variable selection, and classification for Raman algorithm was enhanced using relevance propagation
micro-spectroscopy data. The cross-validated PLS-DA through deep Taylor decomposition, generating heat maps
classification model exhibits remarkable sensitivity and that highlighted important details for each classification.
specificity, scoring 0.94 and 0.85, respectively. Further The study exemplifies the successful integration of MPT

Volume 1 Issue 2 (2024) 51 doi: 10.36922/aih.2775