Predicting RT-PCR test outcomes with machine learning: Guided autoencoder-based imputation and clinical utility assessment

Nowadays, machine learning (ML) is widely used in the healthcare industry for various purposes, including drug development, disease diagnosis and prediction, medical imaging analysis, and outbreak forecasting. Since its emergence in 2019, the COVID-19 pandemic has posed serious challenges to global health. The World Health Organization (WHO) reports that more than 6.9 million people have died from COVID-19, and there have been over 767 million confirmed cases worldwide. Additionally, 936 fatalities and more than 359,000 new cases were reported globally between May 12 and June 8, 2025. These figures underscore the burden on healthcare systems, especially in underdeveloped countries with limited funding. This thesis is organized into three main sections. In the first section, machine learning-based models were developed and evaluated using clinical data to predict the likelihood of active COVID-19 infection and recommend the optimal medical facility for treatment. Since the dataset contains missing data, several imputation techniques are employed. The second section of this study assesses multicollinearity and its impact on the effectiveness of imputation methods, ensuring the reliability of these predictive models. In the third section, an autoencoder-based model with noise-aware masked loss for imputing missing data in high-dimensional datasets is proposed. The proposed approach combines structured noise and a guided loss function to improve imputation accuracy and stability while preserving original data relationships. Collectively, these contributions aim to enhance predictive modeling and robust data completion in healthcare, supporting more trustworthy AI-driven decision-making for current and future public health challenges.