Divya Nagpal

This project aims to leverage audio data from movies to detect and classify various low-level intentions. These predictions are further utilized in generating the motion codes that are sent to the actuators installed in the D-Box theatre seats. In order to perform sound event detection, i.e, detecting the presence and location of the low level intention, we require the audio waveforms and their annotations containing the temporal information. With the help of clustering, we have generated the pseudo labels for the annotations. Next using these clusters as class labels we set up a pipeline for the model architecture FDY-CRNN for this task. Experiments with different possible combinations for layer types: dynamic vs vanilla convolutions, network depth, optimizers, data filtering, learning rate scheduler, etc. were conducted and progressively better results were obtained.

Developed the “Time to Response” (TTR) models with the aim to recover maximum amounts of account receivables (A/R) in as short a time as possible, thereby reducing AR days and maximizing the hospital revenues. These survival models capture the previous response time patterns and predict the probability of getting a response for each claim filed, from the day of claim file submission. The Random Survival Forest model performance is assessed in terms of Concordance Index and Integrated Brier Score (IBS) metric.

Cash Forecasting was developed for billing entities under Adventist Health (AH) for quarterly cash forecasting (predict 90-day future cash inflow), using cash data and calendar variables. Five years of data extract was used for modeling with four models (random forest, VARMAX, ARIMAX, and moving average). Additionally, Django based framework was built for EDA to analyze time series decomposition, trend and seasonality patterns, outlier detection, and lag analysis with features.

Malaria is a significant burden on our healthcare system and it is the major cause of death in many developing countries. Therefore, early testing is necessary to detect malaria and save lives. The main aim of this malaria detection system is to address the challenges in the existing system and reduce the workload of a medically trained professional by automating the process of malaria detection using Deep learning and image processing. A Convolutional Neural Network model was built which achieved an accuracy of 94% and an f1-score of 0.92. Data augmentation was done to make the dataset balanced.