Prediction of Smoking Habits From Class-Imbalanced Saliva Microbiome Data Using Data Augmentation and Machine Learning

Abstract

Human microbiome research is moving from characterization and association studies to translational applications in medical research, clinical diagnostics, and others. One of these applications is the prediction of human traits, where machine learning (ML) methods are often employed, but face practical challenges. Class imbalance in available microbiome data is one of the major problems, which, if unaccounted for, leads to spurious prediction accuracies and limits the classifier's generalization. Here, we investigated the predictability of smoking habits from class-imbalanced saliva microbiome data by combining data augmentation techniques to account for class imbalance with ML methods for prediction. We collected publicly available saliva 16S rRNA gene sequencing data and smoking habit metadata demonstrating a serious class imbalance problem, i.e., 175 current vs. 1,070 non-current smokers. Three data augmentation techniques (synthetic minority over-sampling technique, adaptive synthetic, and tree-based associative data augmentation) were applied together with seven ML methods: logistic regression, k-nearest neighbors, support vector machine with linear and radial kernels, decision trees, random forest, and extreme gradient boosting. K-fold nested cross-validation was used with the different augmented data types and baseline non-augmented data to validate the prediction outcome. Combining data augmentation with ML generally outperformed baseline methods in our dataset. The final prediction model combined tree-based associative data augmentation and support vector machine with linear kernel, and achieved a classification performance expressed as Matthews correlation coefficient of 0.36 and AUC of 0.81. Our method successfully addresses the problem of class imbalance in microbiome data for reliable prediction of smoking habits.

Association of general health and lifestyle factors with the salivary microbiota – Lessons learned from the ADDITION-PRO cohort

“…Nevertheless, it is noteworthy that it was still possible to build a good classifier based on the salivary microbiota to differentiate between current smokers and current non-smokers in the present study. Indeed, we are not the first to do so, as in 2022, another study used data on the salivary microbiota from 175 current smokers and 1070 ex- and non-smokers to create a model that achieved an AUC of 0.81 ( Díez López et al., 2022 ), which is comparable with our findings ( Table 2 ). Importantly, when considering that smoking, which accounted for only 3.2% of differences in our dataset, can deliver such a robust model, this reinforces the concept of using the salivary microbiota for identification of oral diseases, as periodontitis has been shown to have a much higher impact on the composition of the salivary microbiota, than smoking ( Belstrøm et al., 2021 ).…”

Targeted DNA methylation analysis and prediction of smoking habits in blood based on massively parallel sequencing

“…Additionally, smoking habit predictions can be corrected with other traits (such as sex and age) for improved accuracy and combined for a more complete picture of a personalized epigenomic fingerprint. As a final point, efforts to discover and combine other types of molecular biomarkers of smoking, such as single nucleotide polymorphisms [90][91][92], RNA markers [84,93] and microbial DNA [94], are worth exploring.…”

Detection and Classification of Cancer and Other Noncommunicable Diseases Using Neural Network Models

“…Notably, the computer vision field has utilized data augmentation with great success [176][177][178] . Efforts to use neural networks on biological data have seen limited use of data augmentation techniques to improve model performance with research focusing on single cell RNA, methylation, and SNP data [179][180][181][182] . All of these efforts focus on generating new samples from generative models such as variational autoencoders, generative adversarial networks or deep Boltzmann machines.…”

“…All of these efforts focus on generating new samples from generative models such as variational autoencoders, generative adversarial networks or deep Boltzmann machines. They rely on the variable and imperfect nature of the generative process of these models to produce new samples that are somewhat different from their authentic counterparts 179,180,183,184 . This effort is excellent at building datasets that normalize sources of confounding variations such as batch effects but do not address the need for missing samples in the population.…”