(A-6) Support Vector Machine vs Convolutional Neural Network for Classification of Single Cancerous Extracellular Vesicles spectral signals

Introduction:: In recent years the rapid development of data science and machine learning has enabled scientists to use these powerful tools for the analysis of complex and large datasets in many fields. In the context of health and medical care, using machine learning to analyze medical and clinical data can provide solutions to challenges such as early and accurate monitoring of disease onset and progression, as well as personalized treatment. The continuous innovation of biosensors has generated complex heterogeneous biological data in ever-increasing quantities, such as spectral libraries generated via surface-enhanced Raman spectroscopy (SERS). SERS, an optical characterization technique, can be used to collect information on the molecular composition of samples over the area studied. Through SERS, scientists have studied Extracellular vesicles (EVs), nanosized vesicles recently appointed as potential cancer biomarkers since they are rich in information about their cells of origin. However, as many cells produce EVs, their populations found in biofluids of cancerous samples are heterogenic and of an intrinsic complexity. Multiple machine learning algorithms have been used to analyze SERS spectra, like discriminant analysis, k-nearest neighbors, convolutional neural networks (CNN), and support vector machine (SVM). Nonetheless, most studies have utilized spectra collected from multiple EVs. Our in-lab-developed SERS-assisted nanostructured array collects the spectral fingerprints of single EVs via SERS. Here, we present a comparison between SVM and CNN models to determine the model with the highest performance to output information on the biomarker status when the EVs spectral dataset is integrated by single EVs measurements.


Materials and Methods:: We use a single-EVs spectral library previously collected in Mahshid Lab with an in-lab-developed SERS-assisted nanostructured array that enables single EV spectra collection. The library integrates databases from 10 cancerous glioblastoma multiforme (GBM) cell lines, 12 GBM patient samples, and 8 healthy control samples. All the data was preprocessed (baseline subtraction, normalization, smoothing) prior to its integration into their respective databases within the library. The two models tested are an SVM and a ResNet CNN due to their popular implementation in SERS studies. We optimized the algorithms’ hyperparameters and implemented a 5-fold cross-validation to prevent overfitting. First, we study the performance of the models in terms of the sample size by tracking their prediction accuracy as a function of the increment of the sample size. Then we study SVM and CNN performance in two tasks to test the models according to the needs of an EVs study: multi-cell classification and binary cancerous and non-cancerous discrimination. To ensure the models' appropriate training, we use accuracy and loss curves. The models’ performances are compared in both tasks using the evaluating metrics: accuracy, precision, and area under the receiver operating characteristic curve.


Results, Conclusions, and Discussions:: Results

The CNN used consists of a first convolution layer, followed by batch normalization, two residual layers (3 blocks of two convolutional layers, a batch normalization, and an activation function, rectified linear activation function layers), and a final fully connected layer. The models' optimizations were ensured through Bayesian and grid search for SVM and CNN, respectively. The performance of the models for the first task cell line classification was 84% and 70,4% accuracy for CNN and SVM, respectively. For the binary classification between cancerous and non-cancerous, the AUC reached with CNN is 0.89, which is superior to the 0.84 of the SVM.

Discussions

Although deep learning approaches, like CNN, tend to replace classical approaches such as SVM due to their higher accuracy, they need larger data availability and computing power (or cloud platforms) for their appropriate implementation. SVM is a simpler and less demanding model and is still able to generate results in a clinically relevant manner. Data availability is a crucial factor when selecting a model, as for some diseases, it is difficult to acquire a high amount of data due to the rarity of the disease or difficulty in the collection of relevant samples.

Conclusions

The application of machine learning models for diagnostics holds great potential to help overcome the challenge of complex heterogeneous biological data for the successful classification of intricate datasets, such as EVs population in liquid biopsy samples, which could be employed for advancing the technologies toward truly personalized medicine.

Acknowledgements (Optional): :

References (Optional): :