Back to archive

Vol. 1, No. 1 (2022)

Deep Learning and Applications

Deep Learning and its applications in healthcare and finance

A Proposal for a Deep Learning based Credit Scoring approach

Cuong Do, Cory Wang

Interpretation and prediction are two goals of modeling. The goal of interpretation is to extract information from the model about how the response variables are associated to the input variables, while the goal of prediction is to predict what the responses will be given some set of inputs. The dilemma is that interpretable algorithms such as linear regression or logistic regression are often not accurate for prediction, while complex algorithms for better prediction are much more accurate but not easy to interpret1. In this paper, we propose a new credit scoring approach using Deep Learning based on our new Replicated Local Interpretable Model-agnostic Explanations approach. The approach is able to interpret black-box model decisions while provides a better performing model than linear models, so offers a new way for credit scoring for the banking and finance industries.

STL Decomposition-based Stock Price Forecasting using LSTM model

The Bao Pham, Khoi Mai Tieu, Hoang Vuong Pham, Ngoc Thanh Vu, Phan Thu Hang Nguyen

Stock price forecasting is one of the challenging research applications in time series data analysis because it is affected by numerous factors such as political news, weather, and pandemics. In addition, stock prices have several strong statistical univariate properties like unit-root property, the fat tail phenomenon, and volatility clustering. Equally important, time series in the financial market expresses strong seasonality and unpredicted components. To enhance the performance of stock price forecasting, this work proposes the method by combining the seasonal trend decomposition using the Loess (STL) method and long-short-term memory (LSTM) model. The proposed models are experimentally evaluated on 50 stocks collected from Forex, commodity, and New York Stock Exchange (NYSE) stock market. The experimental results show that LSTM-STL-price enhances the forecasting performance compared to single LSTM model.

Application of deep-learning for medical text analysis: a comparison of different types of recurrent neural network

Cuong Do, Lan Vu

We investigate the application of Deep Learning in the healthcare settings to automatically categorize symptoms using patient transcriptions. Several types of networks are investigated, including using stacked unidirectional and bidirectional long short-term memory networks and gated recurrent unit networks, and combinations of these. The approach obtains high accuracy, precision, and recall. Specifically, the stacked LSTM model, stacked bidirectional GRU model, and the stacked bidirectional LSTM model perform equally well in term of accuracy, while that of the stacked GRU model is slightly lower. These well-performing models could be applied for automatic symptom categorization based on patient transcriptions.

A Deep Learning Approach using EfficientNet-based Transfer Learning for Diabetic Retinopathy Diagnosis

Cuong Do, Lan Vu

Detecting and treating diabetic retinopathy early is crucial to preventing blindness among the working-age population worldwide. The aim of the present study is to evaluate the efficacy of deep learning models in diagnosing diabetic retinopathy. The EfficientNets, which are claimed to be more efficient than many current well-performing network architectures, were tested to see how transfer learning performed with color fundus photography as input.

Convolutional Neural Network for lung diagnosis: A study of COVID-19 cases using chest X-ray and CT images

Cuong Do, Lan Vu, Cory Wang, Minh Hoang

This study aims to investigate an automated approach using Convolutional Neural Network (CNN) to efficiently recognize COVID-19 cases vs healthy cases using chest X-ray and CT images. Several models using pre-trained weights, including VGG16, VGG19, InceptionV3, InceptionResNetV2, Xception, DenseNet201, ResNet152V2, and NASNetLarge were investigated. We concluded that the models trained on the X-Ray image dataset outperform the models using the same architectures trained on the CT scan image dataset, and that the VGG16-based model outperforms all the rest of the models trained on the same X-Ray image dataset, and that model decisions could be interpreted for human understanding via local interpretable model-agnostic explanations (LIME). Comparing the models trained on the two types of image sources regarding their cost, an X-Ray based model for COVID-19 would be a lower cost solution to help physicians to be more confident by having a second opinion in assessing patients where there is shortage of test kits, while X-ray devices are widely available.

A Transformer-based Approach for Automated Findings using X-Ray Images

Cuong Do, Bao Pham, Lan Vu

We propose an automated approach for generating findings from X-Ray images based on multi-layer transformer decoders. Several convolutional neural network architectures were investigated for image extraction. Average cosine similarity based on BERT embeddings is selected as the measure of performance evaluation. This automated approach could be useful in situations where there is a lack of expertise in reading medical images, or when a preliminary result or second opinion is needed.