Data Science

In our hospital, the high volume of pediatric emergency room (PEDER) admissions was a serious problem. With our project, we aim to accurately forecast pediatric emergency room admissions in advance. By doing so, we can adjust doctors’ schedules to allocate more staff during peak times. This will reduce patient waiting times while lowering the doctors’ immediate workload, ultimately enhancing the quality of patient care. Approximately four hundred seventy thousand records covering eight years were analyzed for this model. We’re analyzing past visit trends, seasonal fluctuations, local events, and other relevant data that typically influence emergency room visit volumes. This real-time model, which generates daily, weekly, and monthly forecasting reports about the admissions number of patients, helps optimize physician staffing by predicting overcrowding.

In this project, I built a deep-learning model using audio data collected from early preterm neonates in the NICU. Using labels obtained from Echocardiography for the diagnosis of PDA. This project can help in early, non-invasive detection of PDA, leading to timely interventions and better outcomes for preterm neonates. Subsequently, we converted these audio data into spectrograms and classified them using deep learning algorithms as either having PDA or not PDA. Currently, we have reached two hundred patients and the data collection process is ongoing. We will continue expanding our dataset to improve the model’s robustness and reliability.

During my residency, my most significant project involved building a personalized transfusion strategy for PICU patients. We analyzed historical patient data using machine learning algorithms, focusing on variables that impact transfusion requirements, to develop a predictive model. Our model exhibits high performance, as demonstrated by the ROC curve plot. It allowed for more precise and timely transfusions, reducing unnecessary procedures and potential complications, thereby improving patient care.

A deep learning project for segmenting echocardiography (ECHO) parasternal long-axis (PLAX) view images into six primary heart structures using a U-Net-inspired full convolutional network (FCN) architecture.