Curriculum Vitae

PROFESSIONAL EXPERIENCE

St, Jude Medical, Acquired by Abbott Laboratories, Inc. - Fortune 500 medical product manufacturer.

Various titles and roles | Sylmar, CA | December 2017 - Present

Associate Research Fellow, Volwiler Society

Honorary title given to engineers and scientists who have made sustained and significant contributions.

Artificial Intelligence for Cardiac Anomaly Detection

The Assert IQ™ implantable cardiac monitor (ICM) is a small battery powered sensor that gets implanted under a patient’s skin, where it monitors electrical signals for signs of heart abnormalities. If an abnormality is detected, the ICM sends an alert to the patient’s doctor, along with a recording of the electrical activity for confirmation.

Although ICMs last many years and are much more convenient than short-term wearable monitors, they tend to show relatively high false positive (FP) detection rates due to their narrow electrode spacing, limited processing power, and the complexity of the signals being analyzed. This can cause a significant data review burden for clinicians.

To solve this problem, I led the technical development of an artificial intelligence (AI) product that uses a multi-layer convolutional neural network (CNN) to classify ICM heart recordings. With multiple convolution, batch normalization, dropout, and pooling layers, its performance is on par with human experts: maintaining well over 98% sensitivity while reducing false positive detections by over 70%.

Training the CNN posed some unique challenges; given the expertise required to adjudicate ICM recordings, we had a limited number of labeled examples to work with. To overcome this, I developed a series of custom data augnmentation techniques that allowed us to generate a large number of synthetic training examples from a small set of real recordings. These techniques included randomly stretching or shrinking the signal through linear interpolation, adding a small gaussian noise element, randomly shifting in time, and adding targeted training samples based on what the model gave uncertain classifications in earlier runs.

As the model inference needed to run on an existing non-GPU architecture, efficiency and performance were critical. Unlike many electrogram classification models, which are often based on large image classification models and then adapted via transfer learning, this model is comparatively tiny and doesn’t require expensive signal preprocessing. To further improve performance, I implemented a web API that loaded several concurrent model instances into a shared pool in memory, and used a different memory pool for large electrogram object reuse in order to avoid garbage collection. The web API also used SIMD to standardize and preprocess the input data. The result is a product that can classify hundreds of electrograms per minute on a single CPU (far faster than what is needed in the production system), and millions per hour on a single GPU, for internal research purposes.

As an FDA-regulated product, it was also critical that we followed a rigorous software development lifecycle (SDLC) to ensure the safety and efficacy of the AI. I led the effort to establish a process that enabled our data engineers, data scientists, and software engineers to work together effectively, while maintainingful traceability and compliance with FDA regulations. This process involved the use of a separate dedicated Git repository for each vertical, with the output of each engineering step being saved as a permanent artifact for input to the next step.

Languages and technologies used: Python, C#, .NET, Pytorch, Tensorflow, ONNX, ASP.NET Core Web API, Azure, Ubuntu Linux, Azure DevOps, Docker, CUDA.

Research Data Lakehouse

Patients with pacemakers, implantable cardiac defibrillators (ICDs), and similar implanted medical devices are often remotely monitored. This means that they routinely transmit thousands of diagnostic data elements to a central hub. This hub alerts the patients’ doctors of any potential anomalies or issues, and provdes a web portal for them to review the data.

With millions of patients and billions of data elements, this creates a massive amount of data that can be used for research and new product development. However, the data is stored in a variety of formats in highly restricted systems, making it very cumbersome to access and analyze.

To solve this problem, I led the development of a cloud-based data lakehouse. Based on Azure Databricks, it provides a hightly secure and convenient platform for data scientists and researchers to access and analyze the data. It also massively reduces the time and effort required to prepare the data for analysis. Tasks which previously took days or weeks are now routinely completed in minutes or hours. Because of the way in which the data is partitioned and preprocessed, most queries can be run with minimal compute resources, making it very cost effective, at an average cost of less than $20/hour.

One of the key challenges was to reliable de-identify and transfer the data over a slow, unreliable connection. To solve this, I developed a custom Extract, Transform, Load (ETL) pipeline that uses a processing backlog and commit log to ensure that the data is transferred reliably and with minimal duplication. Parallel processing allows it to easily extract a week’s worth of data in only a few minutes.

Overall, this solution has been critical to the successful delivery of several Abbott medical device products.

Languages and technologies used: Python, C#, .NET, Azure, Databricks, Apache Spark, Parquet, Delta tables, Oracle, SQL, Apache Cassandra, RDF.

Ventricular Fibrillation Therapy Assurance

Ventricular fibrillation (VF) is a life-threatening heart arrhythmia that can lead to sudden cardiac arrest. Implantable cardiac defibrillators (ICDs) are designed to detect VF and deliver a shock to restore normal heart rhythm.

Some VF has a polymorphic electrical signature that can be difficult for ICDs to detect reliably. To improve poloymorphic VF detection while minimizing false positive detections, I joined a team of two clinical engineers to develop the Ventricular Fibrillation Therapy Assurance (VFTA) algorithm, which is a key feature in later ICD models. This algorithm works by analyzing the electrical signals from various sensing channels and looking for signs of dropout that are indicative of VF. If droput is detected, the algorithm will reprogram the ICD to used more aggressive VF detection thresholds, which allows it to detect and treat VF more reliably.

My most notable contribution was to implement a platform which was able to quickly merged millions of stored electogram recordings into continuous heart rhythm strips. A naive approach would have been too slow to be practical, but I was able to develop a custom algorithm that looked for overlap in electrogram markers and used that to quickly merge the recordings together. I then implemented a VFTA simulator which could run the algorithm on these continuous strips and generate a large number of synthetic results in only a few seconds. This allowed us to rapidly refine, test, and validate the algorithm.

Languages and technologies used: C#, .NET.

Artificial Intelligence for Lead Noise Detection

Designed and trained a convolutional neural network (CNN) to detect electromagnetic noise in stored electrograms. The resulting model achieved high prediction accuracy comparable to a human expert, and has since been used multiple times to characterize the company’s and competitors’ pacing lead performance.

Languages and technologies used:* Python, C#, .NET, Pytorch, Tensorflow, ONNX, ASP.NET Core Web API, Azure, Ubuntu Linux, Azure DevOps, Docker, CUDA.

Merlin Insight Data Warehouse and Team

Worked with small international team to develop the company’s first data warehousing solution, which subsequently led to the creation of a dedicated team of analysts who used it characterize field issues, characterize performance, and offer guidance in new product development. Critical business decisions that previously were made using tiny ad hoc samples were then able to be much more fully supported. This solution was developed using a combination of an RDF triple store and an Apache Cassandra database running on a multi-server, on-premises Linux server cluster.

Languages and technologies used: C#, .NET, Java, RDF, Apache Cassandra, SPARQL

Rapid Transmission Parser

Abbott medical devices routinely transmit their raw memory contents to a database, where a software process then parses the diagnostic information for communication to clinicians. This software process works fine for a relatively slow incoming transmission rate, but for research and new product developent, where being able to parse select data elements from a backlog of millions of transmissions is essential, this process far too slow.

To solve this problem, I created a parser grammar using Another Tool for Language Recognition (ANTLR). This grammar describes an internal domain specific language that is used to translate raw device memory into meaningful data elements. This mapping is quite challenging due to the many-to-many relationships and complex algebraic equations used in the specification, and with hundreds of thousands of data elements that vary per product, coding the parsing logic by hand would have been impractical.

ANTLR was then used to translate the interface specification into compileable code in C# or Java, where it could then be used to create a parsing library.

The resulting parser achieved over a 10,000-fold increase over the existing software process, thereby enabling the company to conduct analyses in hours or days that previously would have taken months or years. This parser has been a core part of many subsequent product development efforts and urgent field investigations over the years.

Languages and technologies used: C#, Java, ANTLR, compiler generators.

Automated Test Frameworks

I’ve designed and implemented two expansive automated test frameworks that have been reliably used by hundreds of engineers since their inception many years ago. These frameworks solved the following key problems that testers faced at the time:

• Auto-Navigation: I created a self-updating directed acyclic graph (DAG) of programmer UI elemements. When an automated test needs to put the system in a known pre-condition state, it no longer has to be updated every time the UI changes. Instead, the test can make a single function call, which then traverses the DAG and interacts with the system under test to automatically create the necessary precondition.
• Automatic Test Generation: Recognizing that many medical device tests could be represented as state machines, I devised a simple domain-specific event sequence language which allowed testers to describe the scenario in a few hundred characters. My libraries then tokenized this input and generated a state machine which interacted with the test equipment. The result was over a 10X reduction in test script development time.
• Test Results Management: I integrated the test frameworks with an internal test result tracking system, allowing test teams to automatically dispatch automated tests to hundreds of test platforms simultaneously. I also created several sophisticated test log visualization tools which allow testers and developers to easily debug. These tools include memory snapshotting, hierarchical data collapsing, variable monitoring, and more.
• Documentation and Training: To enable verification engineers without formal software development backgrounds to contribute effectively, I created extensive framework wiki documentation and convenient wrapper classes.

Languages and technologies used: C++, C#, CORBA, TCL.