Page 38 - 2023

Ph.D.
(Engineering & Technology)
DEVELOPMENT OF PREDICTIVE ANALYTICS MODEL FOR DISEASE
PREDICTION USING MACHINE LEARNING TECHNIQUES

Ph.D. Scholar : Chauhan Hetal Bhupendrasinh
Research Supervisor : Dr. Kirit J. Modi

Regi. No.: 17276341006
Abstract :
A remarkable amount of research has been proceeding to apply machine learning
techniques to produce healthcare solutions due to availability, adaptability and
advancement of cloud and web technologies. On the other hand, COVID19 diagnosis
process in its current form is facing the problems of shortage of medical resources with
high growth of confirmed cases that results in large waiting time for screening of
COVID19 patients. Increase in diagnosis time enhances the chances of cross infection.
Essential requirement to stop the outbreak is early diagnosis of COVID19 patients. Even
though cases are under control, people living to remote places cannot get timely
treatment due to unavailability of specialized experts. Machine learning based predictive
models can come up with the solution to the issues of COVID19 diagnosis by assisting in
initial screening as well as helping experts in decision making.
Researchers used machine learning classifiers to predict COVID19 positivity. We also
propose a framework to classify a sample with clinically assessed parameters, patient-
reported symptoms, past medical histories into COVID19 positive or negative. Novelty of
our work includes the method to find optimal subset of features for improved
performance of classification. Results of experiments revel that a classifier trained with
features selected by the proposed method got better accuracy of around 3% to 12 %
compared to a classifier trained on all the features in the dataset.

Efforts have been applied to develop COVID19 prediction system using either of chest X-
rays, CT scans or clinical parameters. Few works have been done incorporating multi-
modal inputs. Most approaches developed to predict disease positivity only without
predicting disease severity. Predicting severity of infection is important in resource
management and reducing mortality. We propose reliable multimodal framework to
classify a sample into mild, moderate or severe class denoting infection severity. Sample
includes text data (Patient details, co morbidities, blood parameters) and X-ray images.
We also propose AMSFmap methodology to improve image classification result.
Experiment results concluded that incorporating AMSFmap in existing CNN model
results increase in accuracy of the model to around 4% to 9% and the proposed approach
significantly outperformed existing approaches to classify COVID19 chest X-ray images.

13