Ankit Kumar

Designed, Created and Trained a Self-Supervised learning Model using an ECG and its corresponding EHR sequence as Input, A CNN was used to encode an ECG, and transformer to encode the sequence of EHR. Finally, contrastive loss was used to bring an ECG and its corresponding EHR closer to each other and dissimilar ECG, EHR pairs far part in the latent space. The weights learned for the CNN were used as pre-trained weights for several disease models improving the performance significantly for each disease model. Designed and Created the Data pipeline for ECG Viewer where, given an ECG find all the different types of abnormalities that are present in it and also provide some insights about the ECG – Designed Created and Trained a UNET model which was used to segment an ECG into different waves. Post-segmentation different ECG parameters were computed and using these parameters some basic ECG abnormalities were computed. Finally created APIs using which a user can get the ECG parameters and the basic ECG abnormalities. Designed Created and Trained a multi class model to detect some complex abnormalities present in an ECG. Initially, ECG Reports were used for ground truth labelling an ECG. Created APIs using which a user can get some complex abnormalities that are present in an ECG. Integrated the full Data Generation (The above two bullet points) pipeline using airflow generating all the ECG parameters and abnormalities data for all the ECGs in which a user interested in and saving in MongoDB. The UI team used the APIs to get the ECG parameters and abnormalities to showcase in its app. Designed and Created and Trained an Encoder Decoder Model for the detection of abnormality using an ECG. Initially a CNN model was being used for the detection of abnormality present in an ECG. The decoder was added using LSTM which takes as input the encoded output of CNN and an attention layer was added to all the time steps of LSTM to get the final output. The model eventually improved

Part of the 4G LTE OAM team and full responsibility of multiple blocks. Responsible for understanding the features and customer requirements and perform implementation accordingly.

Objective: CrowdSense, which is going to represent the objects seen in some particular region in a meaningful format. Developed the system, finding the density of different objects in a particular scene post detecting them using yolov3.