t-Distributed Stochastic Neighbor Embedding Method Research Articles

Generating a multimodal artificial intelligence model to differentiate benign and malignant follicular neoplasms of the thyroid: A proof-of-concept study

BackgroundMachine learning has been increasingly used to develop algorithms that can improve medical diagnostics and prognostication and has shown promise in improving the classification of thyroid ultrasound images. This proof-of-concept study aims to develop a multimodal machine-learning model to classify follicular carcinoma from adenoma. MethodsThis is a retrospective study of patients with follicular adenoma or carcinoma at a single institution between 2010 and 2022. Demographics, imaging, and perioperative variables were collected. The region of interest was annotated on ultrasound and used to perform radiomics analysis. Imaging features and clinical variables were then used to create a random forest classifier to predict malignancy. Leave-one-out cross-validation was conducted to evaluate classifier performance using the area under the receiver operating characteristic curve. ResultsPatients with follicular adenomas (n = 7) and carcinomas (n = 11) with complete imaging and perioperative data were included. A total of 910 features were extracted from each image. The t-distributed stochastic neighbor embedding method reduced the dimension to 2 primary represented components. The random forest classifier achieved an area under the receiver operating characteristic curve of 0.76 (clinical only), 0.29 (image only), and 0.79 (multimodal data). ConclusionOur multimodal machine learning model demonstrates promising results in classifying follicular carcinoma from adenoma. This approach can potentially be applied in future studies to generate models for preoperative differentiation of follicular thyroid neoplasms.

Selection of breast features for young women in northwestern China based on the random forest algorithm

In the research of breast morphology, numerous breast features are measured, whereas only a few parameters are adopted for classification. Therefore, how to extract the key variables from the multi-dimensional features in a rational way is an issue that is focused upon. This study aimed to reduce the complexity of the dimensionality reduction for further improving the objectivity and interpretability of the selected breast features. Since the random forest (RF) algorithm can quantify the feature importance during training, the method was adopted to determine the optimal breast features for classification and recognition in this paper. Firstly, the anthropometric data of 360 females from northwestern China aged from 19 to 27 years were measured by non-contact three-dimensional body scanning technology and the contact manual measurement method. Then, the k-means clustering was applied to categorize breast shapes, and the RF algorithm was utilized to quantify and rank the importance of 25 breast features. Finally, to verify the availability of the RF algorithm on breast feature selection, the t-distributed stochastic neighbor embedding method was adopted to visualize the distribution of breast shape clusters into two dimensions. Meanwhile, four neural networks were determined to recognize the breast morphology. The results demonstrate that fewer breast features can effectively increase the accuracy of breast shape classification and recognition. The best performance of breast shape classification and recognition is obtained when the number of breast features is 13. In this case, the average Hamming loss of four neural networks is the smallest (0.1136). Interestingly, the bust circumference and the horizontal curve of breasts across the bust points are found to be the most important of the 25 breast features in this paper. The importance of the breast curve features is higher than that of the breast cross-sectional features, while the breast positioning features have the lowest importance. Meanwhile, the RF algorithm is verified to be more effective than traditional dimensionality reduction methods, such as principal component analysis, hierarchical clustering, and recursive feature elimination. The approach developed in this paper can be generalized to the dimensionality reduction of other body morphology.

Graph Convolutional Network Enabled Two-Stream Learning Architecture for Diabetes Classification based on Flash Glucose Monitoring Data

The classification of type 1 and type 2 diabetes is currently performed based on biochemical indicators and clinical experience. However, considering the unsatisfactory efficiency and accuracy of the experience-based diabetes type classification, we aim to propose a data-driven diabetes classification model through exploiting features contained in flash glucose monitoring (FGM) data. In particular, we propose a novel data reorganization and topologization method to reasonably extract the features of glycemic variability influence. Furthermore, a graph convolutional network is adopted to learn the inter-day influence feature and a Long Short-Term Memory network to characterize intra-day glycemic variability, which enables simultaneous characterization of slow and fast dynamics in FGM data. Finally, to visualize the effectiveness of our model, a t-distributed stochastic neighbor embedding method is implemented. The effectiveness of the proposed model is evaluated through a cross-validation approach using a dataset containing FGM records of 113 diabetic subjects. Compared with classical machine learning algorithms and neural networks, the proposed model achieved the highest specificity value (0.9943) in diabetes type classification, F-Measure (0.8824) and Matthews correlation coefficient score (0.8250). The obtained results indicate the feasibility of achieving diabetes classification by learning the patterns hidden in continuous glucose monitoring data.

T-Distributed Stochastic Neighbor Embedding Method with the Least Information Loss for Macromolecular Simulations.

Dimensionality reduction methods are usually applied on molecular dynamics simulations of macromolecules for analysis and visualization purposes. It is normally desired that suitable dimensionality reduction methods could clearly distinguish functionally important states with different conformations for the systems of interest. However, common dimensionality reduction methods for macromolecules simulations, including predefined order parameters and collective variables (CVs), principal component analysis (PCA), and time-structure based independent component analysis (t-ICA), only have limited success due to significant key structural information loss. Here, we introduced the t-distributed stochastic neighbor embedding (t-SNE) method as a dimensionality reduction method with minimum structural information loss widely used in bioinformatics for analyses of macromolecules, especially biomacromolecules simulations. It is demonstrated that both one-dimensional (1D) and two-dimensional (2D) models of the t-SNE method are superior to distinguish important functional states of a model allosteric protein system for free energy and mechanistic analysis. Projections of the model protein simulations onto 1D and 2D t-SNE surfaces provide both clear visual cues and quantitative information, which is not readily available using other methods, regarding the transition mechanism between two important functional states of this protein.