Imbalanced data preprocessing techniques for machine learning: a systematic mapping study

Acupuncture indication knowledge bases: meridian entity recognition and classification based on ACUBERT

In acupuncture diagnosis and treatment, non-quantitative clinical descriptions have limited the development of standardized treatment methods. This study explores the effectiveness and the reasons for discrepancies in the entity recognition and classification of meridians in acupuncture indication using the Acupuncture Bidirectional Encoder Representations from Transformers (ACUBERT) model. During the research process, we selected 54 593 different entities from 82 acupuncture medical books as the pretraining corpus for medical literature, conducting classification research on Chinese medical literature using the BERT model. Additionally, we employed the support vector machine and Random Forest models as comparative benchmarks and optimized them through parameter tuning, ultimately leading to the development of the ACUBERT model. The results show that the ACUBERT model outperforms other baseline models in classification effectiveness, achieving the best performance at Epoch = 5. The model's "precision," "recall," and F1 scores reached above 0.8. Moreover, our study has a unique feature: it trains the meridian differentiation model based on the eight principles of differentiation and zang-fu differentiation as foundational labels. It establishes an acupuncture-indication knowledge base (ACU-IKD) and ACUBERT model with traditional Chinese medicine characteristics. In summary, the ACUBERT model significantly enhances the classification effectiveness of meridian attribution in the acupuncture indication database and also demonstrates the classification advantages of deep learning methods based on BERT in multi-category, large-scale training sets. Database URL: http://acuai.njucm.edu.cn:8081/#/user/login?tenantUrl=default.

Breast Cancer Molecular Subtype Prediction: A Mammography-Based AI Approach

Breast cancer remains a leading cause of mortality among women, with molecular subtypes significantly influencing prognosis and treatment strategies. Currently, identifying the molecular subtype of cancer requires a biopsy-a specialized, expensive, and time-consuming procedure, often yielding to results that must be supported with additional biopsies due to technique errors or tumor heterogeneity. This study introduces a novel approach for predicting breast cancer molecular subtypes using mammography images and advanced artificial intelligence (AI) methodologies. Using the OPTIMAM imaging database, 1397 images from 660 patients were selected. The pretrained deep learning model ResNet-101 was employed to classify tumors into five subtypes: Luminal A, Luminal B1, Luminal B2, HER2, and Triple Negative. Various classification strategies were studied: binary classifications (one vs. all others, specific combinations) and multi-class classification (evaluating all subtypes simultaneously). To address imbalanced data, strategies like oversampling, undersampling, and data augmentation were explored. Performance was evaluated using accuracy and area under the receiver operating characteristic curve (AUC). Binary classification results showed a maximum average accuracy and AUC of 79.02% and 64.69%, respectively, while multi-class classification achieved an average AUC of 60.62% with oversampling and data augmentation. The most notable binary classification was HER2 vs. non-HER2, with an accuracy of 89.79% and an AUC of 73.31%. Binary classification for specific combinations of subtypes revealed an accuracy of 76.42% for HER2 vs. Luminal A and an AUC of 73.04% for HER2 vs. Luminal B1. These findings highlight the potential of mammography-based AI for non-invasive breast cancer subtype prediction, offering a promising alternative to biopsies and paving the way for personalized treatment plans.

Optimizing Rare Disease Gait Classification through Data Balancing and Generative AI: Insights from Hereditary Cerebellar Ataxia

The interpretability of gait analysis studies in people with rare diseases, such as those with primary hereditary cerebellar ataxia (pwCA), is frequently limited by the small sample sizes and unbalanced datasets. The purpose of this study was to assess the effectiveness of data balancing and generative artificial intelligence (AI) algorithms in generating synthetic data reflecting the actual gait abnormalities of pwCA. Gait data of 30 pwCA (age: 51.6 ± 12.2 years; 13 females, 17 males) and 100 healthy subjects (age: 57.1 ± 10.4; 60 females, 40 males) were collected at the lumbar level with an inertial measurement unit. Subsampling, oversampling, synthetic minority oversampling, generative adversarial networks, and conditional tabular generative adversarial networks (ctGAN) were applied to generate datasets to be input to a random forest classifier. Consistency and explainability metrics were also calculated to assess the coherence of the generated dataset with known gait abnormalities of pwCA. ctGAN significantly improved the classification performance compared with the original dataset and traditional data augmentation methods. ctGAN are effective methods for balancing tabular datasets from populations with rare diseases, owing to their ability to improve diagnostic models with consistent explainability.

APM-YOLOv7 for Small-Target Water-Floating Garbage Detection Based on Multi-Scale Feature Adaptive Weighted Fusion

As affected by limited information and the complex background, the accuracy of small-target water-floating garbage detection is low. To increase the detection accuracy, in this research, a small-target detection method based on APM-YOLOv7 (the improved YOLOv7 with ACanny PConv-ELAN and MGA attention) is proposed. Firstly, the adaptive algorithm ACanny (adaptive Canny) for river channel outline extraction is proposed to extract the river channel information from the complex background, mitigating interference of the complex background and more accurately extracting the features of small-target water-floating garbage. Secondly, the lightweight partial convolution (PConv) is introduced, and the partial convolution-efficient layer aggregation network module (PConv-ELAN) is designed in the YOLOv7 network to improve the feature extraction capability of the model from morphologically variable water-floating garbage. Finally, after analyzing the limitations of the YOLOv7 network in small-target detection, a multi-scale gated attention for adaptive weight allocation (MGA) is put forward, which highlights features of small-target garbage and decreases missed detection probability. The experimental results showed that compared with the benchmark YOLOv7, the detection accuracy in the form of the mean Average Precision (mAP) of APM-YOLOv7 was improved by 7.02%, that of mmAP (mAP0.5:0.95) was improved by 3.91%, and Recall was improved by 11.82%, all of which meet the requirements of high-precision and real-time water-floating garbage detection and provide reliable reference for the intelligent management of water-floating garbage.

Machine-Learning-Based Prediction of 1-Year Arrhythmia Recurrence after Ventricular Tachycardia Ablation in Patients with Structural Heart Disease

BACKGROUND

Ventricular tachycardia (VT) recurrence after catheter ablation remains a concern, emphasizing the need for precise risk assessment. We aimed to use machine learning (ML) to predict 1-month and 1-year VT recurrence following VT ablation.

METHODS

For 337 patients undergoing VT ablation, we collected 31 parameters including medical history, echocardiography, and procedural data. 17 relevant features were included in the ML-based feature selection, which yielded six and five optimal features for 1-month and 1-year recurrence, respectively. We trained several supervised machine learning models using 10-fold cross-validation for each endpoint.

RESULTS

We observed 1-month VT recurrence was observed in 60 (18%) cases and accurately predicted using our model with an area under the receiver operating curve (AUC) of 0.73. Input features used were hemodynamic instability, incessant VT, ICD shock, left ventricular ejection fraction, TAPSE, and non-inducibility of the clinical VT at the end of the procedure. A separate model was trained for 1-year VT recurrence (observed in 117 (35%) cases) with a mean AUC of 0.71. Selected features were hemodynamic instability, the number of inducible VT morphologies, left ventricular systolic diameter, mitral regurgitation, and ICD shock. For both endpoints, a random forest model displayed the highest performance.

CONCLUSIONS

Our ML models effectively predict VT recurrence post-ablation, aiding in identifying high-risk patients and tailoring follow-up strategies.

Machine Learning Models to Predict Future Frailty in Community-Dwelling Middle-Aged and Older Adults: The ELSA Cohort Study

BACKGROUND

Machine learning (ML) models can be used to predict future frailty in the community setting. However, outcome variables for epidemiologic data sets such as frailty usually have an imbalance between categories, that is, there are far fewer individuals classified as frail than as nonfrail, adversely affecting the performance of ML models when predicting the syndrome.

METHODS

A retrospective cohort study with participants (50 years or older) from the English Longitudinal Study of Ageing who were nonfrail at baseline (2008-2009) and reassessed for the frailty phenotype at 4-year follow-up (2012-2013). Social, clinical, and psychosocial baseline predictors were selected to predict frailty at follow-up in ML models (Logistic Regression, Random Forest [RF], Support Vector Machine, Neural Network, K-nearest neighbor, and Naive Bayes classifier).

RESULTS

Of all the 4 378 nonfrail participants at baseline, 347 became frail at follow-up. The proposed combined oversampling and undersampling method to adjust imbalanced data improved the performance of the models, and RF had the best performance, with areas under the receiver-operating characteristic curve and the precision-recall curve of 0.92 and 0.97, respectively, specificity of 0.83, sensitivity of 0.88, and balanced accuracy of 85.5% for balanced data. Age, chair-rise test, household wealth, balance problems, and self-rated health were the most important frailty predictors in most of the models trained with balanced data.

CONCLUSIONS

ML proved useful in identifying individuals who became frail over time, and this result was made possible by balancing the data set. This study highlighted factors that may be useful in the early detection of frailty.