Nikola Rahman

Revolutionizing ECG Analysis with Cutting-Edge Signal Processing and Machine Learning

I developed signal processing and ML pipelines for ECG processing. I worked with a cardiology expert to develop medical-grade algorithms for several types of ECG analyses and arrhythmia detection models. I was one of the first to apply deep learning to ECG signals. The algorithms comply with essential performance requirements found in EN 60601-2-25 and EN 60601-2-47 standards and are used daily in several clinics for the early detection of heart disease.

I designed and implemented cutting-edge machine learning algorithms and neural network layers using the novel Halide programming language. These algorithms and layers were specifically optimized for high processing speed and low power consumption, making them ideal for use in real-time applications. My contributions to the client's deep learning library included three advanced machine learning algorithms, four state-of-the-art convolutional neural network layers, and an algorithm for neural network quantization. These additions significantly enhanced the library's capabilities and demonstrated my expertise in the field of artificial intelligence and deep learning.

I collaborated with a smart home appliances company to improve their CNN model for real-time image segmentation on a Snapdragon 820 platform. Through my efforts, the model achieved a 50x speed-up while maintaining 99% accuracy. This improvement allows the company to use the model for more efficient and effective image analysis in their smart home products.

The project involved collaborating with a chip design company to implement their custom spiking neural network architecture for processing event-based camera data. I developed custom TensorFlow layers to support the implementation and benchmarked the neural net on known event-based datasets. This enabled the client's research team to experiment with novel neural net architectures optimized for their chip architecture, advancing their research in this area.

I developed deep learning models for an automatic topic breakdown in podcast episodes. Working with a team of 20 engineers and data collectors, we labeled and trained our models on 10,000 episodes and achieved human-level performance. Our algorithm enables users to easily navigate topics within podcast episodes and listen to segments of interest.

I developed a medical-grade system for non-invasive recording of fetal ECG from the maternal abdomen. I designed experiments to determine optimal electrode placement for maximum signal capture and quality and created a signal-processing pipeline for estimating maternal and fetal heart rates and uterine activity. This is the first system of its kind to use dry electrodes, making it easy for moms to record fetal ECG without any skin preparation. The system offers higher signal capture and accuracy than traditional methods like CTG and is comparable to similar systems that use adhesive electrodes.

I have developed an explainable AI (XAI) method for a deep-learning computer vision model that classifies diabetic retinopathy. The method provides stakeholders with insights into the model's predictions, allowing them to better understand how and why the model makes its decisions.

I effectively scaled the training of a deep learning classification model from a single on-premise node to a cloud-based 100+ GPU node cluster. By utilizing advanced techniques such as DistributedDataParallel and automatic mixed precision training, I significantly reduced the model's training time from weeks to less than 24 hours. This enabled faster experimentation and iteration, improving model performance and accuracy.

I led the client's small team of machine learning specialists and helped improve their deep learning models for ECG classification. Through my guidance and implementation of more advanced techniques, we were able to reduce the error rate by over 40%. In addition, I provided expertise on FDA compliance and improved their model validation pipelines. Overall, my leadership and technical knowledge allowed for significant progress in the development of their machine-learning models.