The complexity of multivariate data often necessitates the use of dimensionality reduction methods to simplify the information. Dimensionality reduction of high-dimensional data plays a crucial role in downstream tasks such as pattern recognition, classification, and clustering (Kim et al., 2018). Principal component analysis (PCA) is a classic unsupervised dimensionality reduction method that identifies hidden features in data, providing the most significant signals, and is often used in machine learning. (Kim et al., 2018; Chen and Gao, 2016; Ma and Dai, 2011). PCA simplifies complex data and makes the analysis process easier. Essentially, PCA is an “unsupervised” method that analyzes data purely based on its characteristics, without knowing the grouping of each sample. PCA effectively identifies the “main” elements and structures in the data, removes noise and redundancy, reduces the dimensionality of complex data, and reveals the simple structure hidden behind the complex data (Sugimoto et al., 2012). PCA is widely used in MS-based multi-omics data analysis, particularly for data dimensionality and achieving data visualization (Sugimoto et al., 2012). The results are often visualized using scatter plots.

PCA is commonly applied in clinical analysis to reveal differences between samples, with the distance between samples on the horizontal and vertical axes representing the similarity distance of the samples under the influence of the principal components (PC1 and PC2).

For example, in a single-cell proteomics study of hepatocytes by Rosenberger et al. (2023), PCA was used to reduce the dimensionality of proteomics data, resulting in clear hepatocyte partitioning and demonstrating the biological validity of the data. Similarly, Xu et al. (2022) applied PCA to validate the metabolic profile of mouse liver tissue in a study on the mechanism of action of Huang Qin decoction for treating diabetic liver injury.

PCA is particularly suitable for initial exploratory data analysis, especially when linear relationships are presumed in the data and the interpretability of components is crucial (Beattie and Esmonde-White, 2021; Ivosev et al., 2008). It is also preferred in scenarios where computational efficiency is a priority. However, PCA may not adequately capture complex nonlinear interactions present in biological data.

t-distributed Stochastic Neighbor Embedding (t-SNE) is a nonlinear dimensionality reduction technique primarily designed for data visualization and excels at identifying and discovering complex nonlinear structures in data (Cieslak et al., 2020). It converts high-dimensional data into a low-dimensional space, typically two or three dimensions while preserving the local structure of the data (Cieslak et al., 2020; Da Silva Lopes et al., 2020). t-SNE focuses on preserving the relative distances between similar data points, making it effective for revealing clusters and patterns (Chatzimparmpas et al., 2020).

In the study by Liang et al. (2022), 1,681 proteins were analyzed through proteomics in 258 HCM patients. t-SNE was utilized to visualize and reduce the dimensionality of the data, revealing that the four molecular subtypes were well separated.

t-SNE is ideal for visualizing small to medium-sized datasets where the primary goal is to understand and explore local structures and clustering. It is particularly suitable for high-dimensional data with complex, nonlinear relationships (Gisbrecht et al., 2015). However, it is relatively slow, especially when applied to large datasets.

Uniform manifold approximation and projection (UMAP), an algorithm developed by McInnes et al., is a nonlinear dimensionality reduction technique (Lundberg and Lee, 2017). It is another nonlinear dimensionality reduction method particularly effective at preserving both local and some global structures of the data (Becht et al., 2018; Yang et al., 2021). UMAP is based on manifold learning techniques and constructs a high-dimensional graph representation of the data, which is subsequently optimized to create a low-dimensional embedding (McInnes et al., 2018).

Like t-SNE, UMAP is also effective in capturing complex nonlinear relationships in the data. In general, UMAP is faster and more scalable than t-SNE and it better preserves both local and global structures.

UMAP is preferred for large datasets requiring a balance between local and global structure preservation (Sainburg et al., 2021). It is effective for visualizing complex data structures and works well for high-dimensional data (Becht et al., 2018). Compared to PCA, this method is more complex and can be challenging to understand and debug. Although UMAP is faster than t-SNE, it still requires significant computational resources for very large datasets (Roca et al., 2023).

A crucial component of unsupervised learning is the clustering algorithm. Traditionally, cluster analysis is classified as unsupervised learning because it does not involve class labels or quantitative response variables, which are characteristic of supervised learning methods such as classification and regression (Pan et al., 2013). Cluster analysis is a series of different algorithms that divide observation data into different categories or clusters based on distance functions (Blekherman et al., 2011). The goal is to partition the data into groups such that the distance between samples within each group is smaller than the distance between samples in different groups (Blekherman et al., 2011). Hierarchical clustering (HCL) is a clustering method frequently used in marker screening and enables visualization of gene, protein, and metabolite features (Picard et al., 2023; Granato et al., 2018). The results of HCL are commonly visualized using heatmaps.

In biomedical informatics, HCL is often applied to cluster protein sequence data. Proteins with similar structures also have similar functions. Proteins with similar functions can be grouped into categories through clustering, aiding in the study of protein functions. In clinical analysis, hierarchical clustering is utilized to intuitively display relationships between groups and highlights expression differences of characteristic substances.

In the study of MS-based urine proteomics of gastric lesions by Fan et al. (2022), HCL was used to partition 139 differential proteins with VIP>1 into six clusters, revealing dynamic changes from precancerous lesions dynamic changes in gastric cancer.

In an MS-based metabolomics study of cancer cell lines by Li et al. (2019), HCL was employed to assess metabolic similarities between cell lines.

For MS-based spatial proteomics, dimensionality reduction and clustering methods such as PCA, t-SNE and HCL are effective for quality control of MS-based spatial proteomic data and for examining organelle separation (Huang et al., 2022; Karimpour-Fard et al., 2015; Mou et al., 2022; Ringner, 2008).

Partial least squares discriminant analysis (PLS-DA) is also a dimensionality reduction algorithm. Unlike PCA, PLS is a “supervised” mode of partial least squares analysis, meaning that the grouping relationships of the samples are known, allowing for better selection of characteristic variables that distinguish each group and determination of relationships between the samples. DA stands for discriminant analysis. PLS-DA employs the partial least squares regression method to “reduce the dimensionality” of the data, establish a regression model, and conduct discriminant analysis on the regression results. PLS-DA is particularly suitable for selecting and interpreting metabolite signatures when studying biological systems (D'Andrea et al., 2023).

In an MS-based urine proteomics study of gastric lesions by Fan et al. (2022), PLS-DA was used to analyze the proteomics data of different groups and screen out 139 differential proteins with VIP>1.

In the study by D'Andrea et al. (2023), the PLS-DA model was confirmed through cross-validation, and the average variable importance in projection (VIP) score was used to identify metabolites that differed among sample classes.

Decision trees (DTs) employ ML techniques to address classification and prediction problems. Nodes and leaves are the primary elements that form a decision tree (Chaubey et al., 2020). Nodes test specific properties, and leaves represent a class (Mesarić and Šebalj, 2016). Common decision trees include the ID3 tree, the C4.5 tree (information gain rate), and the CART tree (Gini coefficient) (Ross, 1993; Leo et al., 1984).

The ID3 algorithm is one of the classic decision tree algorithms. The C4.5 algorithm is an improvement upon the ID3 algorithm and can handle discontinuous features (Navada et al., 2011). The ID3 and the C4.5 algorithm are primarily used to address classification problems, but cannot be used to apply regression problems (Singh and Giri, 2014). The CART algorithm can manage both classification and regression problems.

Fannes et al. (2013) introduced CP-DT (Decision Tree Cleavage Prediction), an algorithm based on an ensemble of decision trees trained on publicly available peptide identification data from the PRIDE database. The study demonstrated that CP-DT can accurately predict trypsin cleavage (Fannes et al., 2013).

Decision tree algorithms are fast, however, they are generally not as accurate as other models.

Random forest (RF) is a regression tree technique that employs bootstrap aggregation and predictor randomization to achieve a degree of predictive accuracy (Steven, 2017). Proposed by Breiman in 2001, RF employs randomization to create numerous decision trees and is a widely used tool for classification and regression in bioinformatics and related fields (Steven, 2017; Janitza et al., 2016). Compared to a single decision tree, a random forest exhibits stronger generalization performance. In classification problems, the outputs of these decision trees are voted and aggregated into one output; in regression problems, they are averaged and aggregated into one output (Steven, 2017). RF classification is a widely used supervised learning method for developing predictive models in many research settings (Speiser et al., 2019).

The random forest algorithm is simple and easy to implement, applicable to both classification and regression problems (Steven, 2017). It has the following features.

a. It can handle numerous input variables, and the more data features present, the more stable the model (Belgiu and Drăguţ, 2016).

Khalilia et al. (2011) utilized National Inpatient Sample (NIS) data from the Healthcare Cost and Utilization Project (HCUP) to train RF classifiers for predicting eight disease categories. The results demonstrated good performance (Khalilia et al., 2011).

However, RF does not perform as well for regression problems as it does for classification and may not produce good classification results for small or low-dimensional datasets (datasets with fewer features).

Support vector machine (SVM) is a supervised algorithm that learns from examples to assign labels to objects (Boser et al., 1992). Compared to other ML methods, SVM is highly effective at identifying subtle patterns in complex data sets (Aruna and SP, 2011). The purpose of SVM is to create a decision boundary between two categories, facilitating the prediction of a label based on one or more feature vectors (WS, 2006). This decision boundary, called a hyperplane, should be oriented as far away as possible from the nearest data point for each class, referred to a support vector (Huang et al., 2018).

The computational complexity of SVM depends on the number of support vectors rather than the dimension of the sample space, thereby avoiding the “curse of dimensionality” (Markowetz, 2001). However, SVM is sensitive to missing data and is difficult for solving multi-classification problems (Cervantes et al., 2020). In areas where SVM performs poorly, researchers have developed other applications such as SVM for large datasets, multiple classifications, and imbalanced datasets (Cervantes et al., 2020).

Mavrogeorgis et al. (2023) obtained urine peptide data of 1850 healthy controls (HC) and CKD (diabetic nephropathy-DKD, IgA nephropathy-IgAN, vasculitis) participants from the Human Urine Proteome Database. UMAP was combined with SVM for binary (DKD, HC) and multi-class (DKD, HC, IgAN, vasculitis) classification.

K-Nearest Neighbor (k-NN) is a simple and practical supervised learning algorithm frequently used to deal with classification problems (Boateng et al., 2020). It examines the k nearest sample points closest to the new sample point in the training set, using a specific distance metric, and classifies the new sample point into the category with the most occurrences among the k sample points (Abu Alfeilat et al., 2019). The parameter k is crucial, and its value should be optimally chosen (Zhang et al., 2017a). A value that is too low will increase the error rate, while a value that is too high can render the model ineffective (Zhang et al., 2017b).

The algorithm is simple in principle, easy to understand and implement, applicable to multi-classification problems, and requires no additional processing (Chaubey et al., 2020). However, k-NN involves substantial computational effort and requires considerable memory resources. Its performance is influenced by the parameter k and it tends to perform poorly on unbalanced datasets.

eXtreme Gradient Boosting (XGBoost) is a machine learning model built on a decision tree ensemble and is among the most widely used machine learning algorithms (Kavzoglu and Teke, 2022). The algorithm has the following features:

a. It excels in processing both structured and unstructured data, frequently achieving higher accuracy compared to other algorithms (Arif Ali et al., 2023).

Li et al. (2024) developed a model incorporating 17 feature variables using XGBoost, based on the multidimensional data from a retrospective cohort of 274 papillary thyroid carcinoma (PTC) patients. This model demonstrated strong predictive performance in differentiating between low-risk and medium/high-risk PTC cases and was designated as the PTC Preoperative Risk Assessment Classifier (PRAC-PTC).

However, it is sensitive to parameters settings, with the choice of parameters significantly influencing the results (Demir and Şahin, 2022). In some cases, XGBoost may be overly complex and prone to overfitting the training data.

Missing data refers to the situation where data is incomplete due to some reasons during the process of data collection, transmission, and processing (Du et al., 2020). It is a common problem in MS-based omics data analysis (Huang et al., 2023). The simplest way to deal with missing values is to remove samples with missing values. However, if there are many missing values, such as the missing data of LC-MS-based omics data may be in the range of 30%–50%, a large number of samples will be eliminated, resulting in the loss of more useful information (Liebal et al., 2020).

Imputation methods provide an alternative way of handling missing data rather than discarding missing values and associated data (Huang et al., 2023). The mean, median, mode, etc., of the feature can be used to fill the missing values (Emmanuel et al., 2021). However, these simple methods do not consider the relationship between data variables, which sometimes makes the results of data analysis unreliable. Among the methods for dealing with missing values, many other filling methods consider the relationship between data variables (Baraldi and Enders, 2010). ML algorithms, such as regression, k-NN, and RF, can help resolve missing data problems in multi-omics datasets by inferring values based on observed patterns in existing data (Emmanuel et al., 2021; Mirza et al., 2019).

MS-based multi-omics data may have multiple layers of variables and a large number of attributes, so-called high-dimensional data (Arjmand et al., 2022). While high-dimensional data will cause great trouble for subsequent data processing (Cao and Lin, 2015), dimensionality reduction is a crucial step (Fanaee and Thoresen, 2019). It aims to reduce the number of variables considered, making the data more manageable and easier to analyze while retaining as much information as possible. Before applying dimensionality reduction, multi-omics data need to be preprocessed, such as normalization and missing value filling, to ensure that the data is in a form suitable for analysis (Reska et al., 2021). Dimensionality reduction improves computational efficiency, reduces noise while retaining important information, facilitating data processing (Alhassan and Wan Zainon, 2021). Many ML algorithms can facilitate data processing by reducing data dimensionality while retaining important information, such as PCA, t-SNE, PLS-DA, and UMAP.

Clustering is an unsupervised learning method that groups data based on the attributes of the input features (Reel et al., 2021). Classification is a supervised learning method that provides predicted output as a discrete class (Reel et al., 2021). ML algorithms can group samples into clusters or classify them based on distinct patterns present in multi-omics data which can discover subtypes or stratify patients and identify similarities among clustered patients (Goecks et al., 2020).

Feature selection is a key step in multi-omics analysis and helps reduce data dimensionality. In this sense, feature selection has similar motivations to dimensionality reduction as described above. Feature selection can remove irrelevant features and reduce the number of features used in the analysis, thereby reducing the difficulty of the learning task (Li et al., 2017a). It should be noted that the feature selection process must ensure that no important features are lost.

For a multi-omics dataset, a set of attributes is included, some of them may be critical and useful, while others may be useless. Attributes are called features, those that are useful for the current learning task are called relevant features, and those that are useless are called irrelevant features (Kotsiantis, 2011). The process of selecting a subset of relevant features from a given set of features is called feature selection (Jović et al., 2015).

Feature selection is an important data preprocessing process. In real machine learning tasks, feature selection is usually performed after obtaining multi-omics data, and then the learner is trained.

In practical applications, feature selection methods are mainly divided into filter, wrapper, and embedded methods (Venkatesh and Anuradha, 2019).

a. Filter Selection

Filter selection first selects features from the data set and then trains the learner (Li et al., 2017a). The feature selection process is independent of the subsequent learner. It is the simplest and most commonly used method to implement feature selection. The core of the filtering method selection is to sort the features according to their value, to achieve the selection or elimination of any proportion/quantity of features.

Filter selection is computationally efficient and relatively simple to implement, but it ignores the interactions between features (Jiliang et al., 2014).

b. Wrapper Selection

Unlike filter-based feature selection, which does not consider subsequent learners, wrapper selection directly uses the performance of the most important learner as the evaluation criterion for the feature subset (Wald et al., 2013; Cadenas et al., 2013).

Since wrapper selection directly optimizes a given learner, it is better than filter selection in terms of the final learner performance (Cadenas et al., 2013). However, since the learner needs to be trained multiple times during the feature selection process, it requires a huge amount of computation.

c. Embedded selection

Embedded selection integrates the feature selection process with the learner training process, that is, feature selection is automatically performed during the learner training process (Li et al., 2017b). The most commonly used are tree models and a series of ensemble algorithms based on tree models because the model provides important information about feature importance.

Embedded selection combines the advantages of filter selection and wrapper selection in terms of computational cost and performance, but it cannot identify highly relevant features (Remeseiro and Bolon-Canedo, 2019).

The classification model is the most commonly used type of ML models in clinical researches which can output the type of thyroid disease it predicts based on the input features, such as clinical factors and omics data, thereby providing insights into clinical diagnosis (Figure 4). For thyroid disease prediction, ML methods have been applied in various existing research works. Prediction of thyroid disease at its early stages and categorization into cancer or other thyroid disease is very helpful for treating and recovering the maximum number of patients (Gupta et al., 2024). MS-based multi-omics have become a powerful technique for biomarker discovery which is significant for clinic (Torun et al., 2022). However, screening biomarkers from complex MS data requires reliable bioinformatics tools and ML can be a good choice (Torun et al., 2022). Analyzing large and complex datasets by ML models enables the identification of subtle biomarkers and disease signatures, which leads to earlier and more accurate diagnoses (Ng et al., 2023). This is particularly beneficial in complex diseases specially cancer, where early detection can significantly improve prognosis. The integration of ML and MS-based multi-omics not only improves the efficiency of biomarker discovery but also helps develop more accurate and reliable diagnostic tools, ultimately advancing the field of precision medicine and improving patient outcomes (Johnson et al., 2021; Krishnan et al., 2023).

From the perspective of classification, existing researches mainly focused on binary classification problems in thyroid disease classification based on ML models. Kumari et al. (2024) used age, gender, and hormone levels as features and combined them with 3 ML models to classify hyperthyroidism and hypothyroidism. Results showed that XGBoost was the top-performing model for this task (Kumari et al., 2024).

Al-muwaffaq and Bozkus (2016) developed the Machine Learning tool for Thyroid Disease Diagnosis (MLTDD) mainly focused on thyroid gland medical diseases caused by underactive or overactive thyroid glands. The prediction accuracy was in range between 98.7% and 99.8% for testing. MLTDD can effectively help to make the right clinical decision.

Xi et al. (2022) constructed 6 ML models to predict the malignancy of thyroid nodules. RF and Gradient Boosting Machine (GBM) showed better overall diagnostic accuracy and ability to identify malignant nodules. Their method can be used as additional evidence in the preoperative diagnosis of thyroid cancer.

Guo et al. (2022b) constructed a diagnostic model of benign and malignant thyroid tumors. The benign group contained 5 thyroid diseases and the malignant group contained six thyroid diseases. Clinical factors were used as features and RF, XGBoost, LightGBM, and AdaBoost models were constructed. RF model showed the best performance. Their research proposed a model incorporating novel biomarkers which could be a powerful and promising tool for predicting benign and malignant thyroid tumors (Guo et al., 2022b).

Recently, some researches have begun to focus on multi-classes classification problems. Chaganti et al. (2022) used ML models to predict five thyroid diseases. Results showed that the extra tree classifier-based selected feature yields the best results with 0.99 accuracy and an F1 score when used with the RF classifier (Chaganti et al., 2022).

Gupta et al. (2024) used ML models to predict ten thyroid diseases. Results suggested that an accuracy of 0.998 can be obtained using the optimized AdaBoost model by differential evolution (Gupta et al., 2024). These researches applied ML models to a variety of thyroid diseases, which helped to diagnose diseases more conveniently and comprehensively in clinical practice.

ML models can not only be used to classify the types of thyroid disease, but also to determine whether there is a certain gene mutation in the disease. Kwon et al. (2020) used ML models combined with radiomics to predict the presence or absence of B-Raf proto-oncogene, serine/threonine kinase (BRAF) mutation in PTC and the results showed that the classification accuracy of these models was higher than 60%. Although this research provided a new perspective of application of ML in thyroid disease which focused on the gene mutation, the classification model did not show excellent performance in predicting the presence of BRAF mutation in PTC and need to be further validated in a larger dataset to better assess their potential clinical use.

In addition, ML models can also be used to predict treatment trends and prognosis. Aversano et al. (2021) combined ten different ML models with parameters related to the person being treated to predict whether the LT4-based treatment needs to be increased or decreased for the patients with hypothyroidism. This study provides reference insights for clinical treatment plans.

Wang et al. (2024a) presented ML models based on comprehensive predictors to predict the structural recurrence risk of PTC patients. All the patients were treated with thyroid surgery and radioiodine. Twenty-nine perioperative variables consisting of four dimensions (demographic characteristics and comorbidities, tumor-related variables, lymph node-related variables, and metabolic and inflammatory markers) were analyzed (Wang et al., 2024a). The results showed that the RF model achieved the expected prediction effect, with good discrimination, calibration, and interpretability, and revealed the potential of ML models in improving the accuracy of risk stratification of PTC patients (Wang et al., 2024a).

It is also an important task for ML models to classify clinical patient samples based on MS-based multi-omics data. In many cases, it can be used to determine which ion m/z values contribute most to the models, thereby facilitating the identification of biomarkers and further facilitating the diagnosis and classification of thyroid disease (Beck et al., 2024). Recently, an increasing number of researches have focused on combining ML with MS data for thyroid disease classification, even for further treatment strategies, which have provided broader prospect.

Sun et al. (2022) proposed the first protein-based neural network classifier for thyroid nodules. Their research was the first to establish a deep proteome data repository for various thyroid lesions, analyzed a larger sample size and obtained deeper proteome coverage (Sun et al., 2022). The large-scale thyroid proteome map combined with the neural network model demonstrated the power of the classifier and is expected to be quickly applied to clinical practice to supplement the deficiencies of traditional cytopathology (Sun et al., 2022). Their recent work (Sun et al., 2024) established a protein-based model with targeted MS for the diagnosis of FTC and FTA. This model used 24 proteins filtered by XGBoost as features and performed better than gene-based model. The protein model has 95.7% negative predictive value for ruling out malignant nodules (Sun et al., 2024).

Wang et al. (2022b) develop a rapid classification method by ML and MS-based metabolomics to diagnose PTC. The metabolomics of frozen samples were performed by probe electrospray ionization (PESI) mass and SVM and RF models were used. For the classification of PTC from PTC adjacent tissues, SVM performed better than RF. Their another work developed a rapid method to classify the malignant and benign thyroid nodules by PESI-MS-based machine learning (Wang et al., 2022c). For each FNAB sample, only 10 min is needed to determine its malignancy, which is much easier and faster than traditional diagnosis (Wang et al., 2022c).

Chen et al. (2024) explored the potential of rapid thyroid disease screening using the ZrMOF/Au-assisted LDI-MS platform, enabling rapid screening of malignant thyroid disease from benign patients. The authors constructed a panel of 43 key metabolites as features for ML models to discriminate thyroid cancer from thyroid nodules and NN, RF, LR, and SVM models were used for classification (Chen et al., 2024). The results showed that NN had the best classification performance.

Zhu et al. (2024) combined ML with MS-based proteomics and selected four proteins as features with the highest contributions to predict the efficacy of iodine therapy. This research showed the ability to pre-identify PTC patients who are resistant to radioactive iodine therapy (Zhu et al., 2024).

The above articles were retrieved by entering “machine learning” and “thyroid disease” in Google Scholar and some representative ones involving “application of ML in diagnosis of thyroid disease” were selected by us. We showed them in Table 1 in the order we mentioned in this section with three summary measures, sample types, features, and supervised ML models.

By analyzing the above research works, we found that patients’ personal data and hormone parameters are often used as features, such as age, sex, thyroid-stimulating hormone (TSH), total serum triiodothyronine (T3), thyroid binding globulin (TBG) and total serum thyroxin (T4). Among these features, almost every researcher has selected some features for thyroid disease diagnosis work. MS-based proteomics and metabolomics data are also often used but almost only single type is used. In order to analyze thyroid disease more comprehensively, combining metabolomics and proteomics or combining them with other omics data such as genomics and clinical indicators should be further considered in future studies.

A significant advantage of ML in the clinic is its ability to facilitate early intervention (Adlung et al., 2021). By using advanced algorithms and predictive models, ML algorithms can detect potential health risks at an early stage, and effectively classify diseases into different risk categories, which offer a nuanced understanding of the likelihood of patients developing specific health conditions and allowing clinicians to intervene promptly and implement targeted treatments (Beaulieu-Jones et al., 2021; Choudhury and Asan, 2020; Sun et al., 2023). This goes beyond traditional diagnostic methods, providing a more personalized and proactive approach to healthcare (Yadav, 2024).

Saima Sharleen Islam (Sun et al., 2023) et al. used 11 ML models to predict thyroid risk and used accuracy and recall as evaluation indicators. The results show that the ANN classifier outperforms the others in terms of accuracy. Zhao et al. (2022a) et al. used 5 ML models to predict the risk of nodular thyroid disease in coal miners, with the XGB model having the best overall predictive performance.

Wang et al. (2022a) presented ML models based on comprehensive predictors to predict the structural recurrence risk of PTC patients. All the patients were treated with thyroid surgery and radioiodine. Twenty-nine perioperative variables consisting of four dimensions (demographic characteristics and comorbidities, tumor-related variables, lymph node-related variables, and metabolic and inflammatory markers) were analyzed Wang et al. (2022a). The results showed that the RF model achieved the expected prediction effect, with good discrimination, calibration, and interpretability, and revealed the potential of ML models in improving the accuracy of risk stratification of PTC patients Wang et al. (2022a).

MS-based multi-omics data can also be combined with ML for risk stratification in thyroid disease. Li et al. (2024) first reported a Preoperative Risk Assessment Classifier for PTC (PRAC-PTC) which constructed by ML models used clinical indicators, immune indices, genetic feature, and MS-based proteomics as multidimensional features. The results showed that six proteins (DPP7, PDLIM3, Col12A1, CTSL, TUBB2A, and ITGB5) were identified as the best discriminable proteins between low-risk and intermediate-risk/high-risk PTCs (Li et al., 2024). XGBoost showed the best performance among these ML models authors used. PRAC-PTC can increase the accuracy of the preoperative risk stratification and decrease unnecessary surgery or overtreatment.

Wang et al. (2024b) used different ML algorithms to explore the relationship between mixed-semi-volatile organic compounds (SVOCs) exposure and thyroid nodule. The data was collected by GC-MS/MS. RF and AdaBoost models were selected to screen out the features based on their contribution to the models. Weighted quantile sum (WQS) regression and Bayesian kernel machine regression (BKMR) were used to assess the mixed effects of the SVOCs exposure on thyroid nodule (Wang et al., 2024b). The results showed that high levels of exposure to SVOCs increase the risk of PTC and nodular goiters (NG), with Fluazifop-butyl and Fenpropathrin playing a major role (Wang et al., 2024b).

Wang et al. (2024c) proposed a ML-based objective method to individual to predict the risk of pediatric papillary thyroid carcinomas (PPTCs). They collected the clinical factors and MS-based proteomics data and nineteen proteins were selected by ML models to construct a protein-based personalized prognostic prediction model which can stratify PPTC patients into high- or low-recurrence risk groups and provide a suggestion for clinical decision-making and individualized treatment.

The combination of ML and MS-based multi-omics data in thyroid disease not only enhances disease risk assessment but also improves the approach to patient care. Through early intervention, patient stratification, and the implementation of targeted preventive measures, ML makes a significant contribution to improving patient health and building a more personalized and efficient clinical treatment system (Khalifa and Albadawy, 2024; Khalifa et al., 2024).