As the world transitions to machine learning-powered products and solutions, many subject matter experts will need to develop new skills and ways of thinking/operating. In a field as complicated as healthcare, equipping your workforce with a basic knowledge of machine learning (ML) will typically have better results than bringing in an external “AI-/ML expert” who knows little about the business/industry. This upskilling can prove challenging: what are the tools and resources needed to achieve this, and what types of skills are needed and when?
At Nanowear, we ran into this very issue. Our founding team of seven were grounded in biomedical, electrical, and chemical engineering and this talent pool invented and scaled production of Nanowear’s core intellectual property, cloth-based nanosensors. Nanowear sensors are capable of collecting and transmitting electrical, hemodynamic, audio, and biochemical signals directly from the skin. An ML platform that drives positive health outcomes is what makes a platform like ours a compelling business rather than another novel sensor.
Why this matters now
Nearly all products, from implantable devices, to patient-centric platforms, to smart-beds, have one common thread in 2018 and beyond — they are producing petabytes of data in a digital format that can be collected and processed faster than in any period in human history. But even with unprecedented data, there are many organizational challenges that come with building an ML solution to a healthcare challenge. A particularly important question we have been grappling with is: what is the optimal timing to deploy machine learning infrastructure and talent when it is needed in the company’s platform development journey?
Medical products require scientific and engineering expertise to prototype and scale, while simultaneously navigating strict, yet necessary, regulatory hurdles on the path to market. But once the product platform is ready to collect, transmit, and analyze data in the clinical field, where does a team of biochemists and biomedical/electrical engineers turn to when the needs of the company are much more focused on machine learning?
As Nanowear approached this transition, we had a lot of questions: what is the right way for an organization to transition skill sets to focus much more on algorithms and data? At what point does a medtech company secure the resources for data scientists? Should the data science team be involved before the tools for data collection (hardware, sensors) are even built? How big of a team do you need? Will investors trust that you know how to plan deployment of data science resources if you don’t have a data science background?
Nanowear’s data flow pipeline and potential areas to deploy ML/data science infrastructure/hires
Fortunately for Nanowear, we met the Google Launchpad Studio team at the JP Morgan Healthcare conference last January. We knew that machine learning was perhaps the most valuable piece in our product offering that we had not yet tackled but would be critical in leading up to our 500 patient clinical trial in early 2019. We simply were not experienced enough to answer the multitude of planning-related questions directly inherent to calendar budget and resource forecasting.
Operation and Resource Planning for Applied ML
We learned that ML needs to be deployed as far upstream in the infrastructure as possible, but does not necessitate a full-time data science team. The most critical step was preparing our platform for machine learning as early as possible. For Nanowear, this did not mean gutting existing infrastructure, it meant complementing existing infrastructure and amending software quality management standards of traceability to our stack.
We added infrastructure to ensure that we adequately labeled and characterized each channel input of data across varying metrics (electrical, hemodynamic, sound signals, etc.) before moving the data into the cloud. We also built tools that ensured there was a consistent baseline of denoising, since these would be critical paths for any ML-based model. As we started to put this ML-centric infrastructure in place, we formalized our operational and resource planning around the following areas:
Infrastructure for prototyping — For the Nanowear platform, the first critical focus was using data visualization tools, which gave us immediate feedback on how data was being collected in the system. With wearable devices, it’s not uncommon to have noisy data, and the first requirement of a working model is to be able to tell the difference between signal and noise. We found that Dataflow, BigQuery and Data Studio were the most effective tools for data visualization.
Infrastructure for scale — Infrastructure needs dramatically change when focusing on scale as opposed to prototyped solutions. While many cloud based services are relatively pay-as-you-go, there are some larger decisions that impact architecture at a more fundamental level. For healthcare companies in particular, there are several regulatory requirements on data security and compliance that need to be in place before collecting patient data. While this may be somewhat nuanced, it is absolutely essential infrastructure component for a healthcare company like Nanowear.
Machine learning talent — Ahead of hiring a full-time team for ML, we are focused on leveling up our engineers with ML skills. Especially in healthcare, which rightfully has many rules and regulations, domain experience is a huge asset. As Nanowear grows to be a platform which tries to change health outcomes in many different areas, we will have more need for a dedicated ML/data science team. Until then, the focus is on our first clinical trial — improving outcomes for heart failure patients.
Organizational Resource Planning — For early-stage companies, there isn’t much long-term forecasting we can do regarding fully scaled system cost, since we don’t know exactly what the needs will be of the fully deployed system. What we do know, however, are the specific needs for the upcoming trials: how many patients, how much data, and trial duration. We have taken an approach that seeks to keep the infrastructure as modular as possible as we incrementally scale, while focusing on accurate budgeting and planning for the upcoming trial.
It is our belief that most companies of the future will employ some form of ML. This transition is one of the greatest challenges and opportunities for businesses across domains, but it is particularly relevant for healthcare companies. As alumni of the Google Launchpad Studio program, we are far more confident in understanding our needs, resources, and timing in becoming a machine learning healthcare company. We hope this blog can be your first step in building an operational plan for integrating ML into your business.
Based on proprietary cloth-based nanosensor technology, Nanowear is a New York-based connected-self technology and digital therapeutic platform for diagnostic and chronic disease management and is the first-and-only company in the world to have received FDA 510(k) clearance for cloth-based nanosensor cardiac remote monitoring.
Nanowear improves patient outcomes through a Congestive Heart Failure (CVF) management tools by capturing physiological information, and getting physicians higher-resolution information as to what is happening with the patient. Nanowear’s technology platform has applications across multiple markets such as pharmaceutical clinical research companion devices, in-patient monitoring, risk mitigation monitoring of first responders (EMTs, Firefighters, etc), and industrial safety workers.
Nanowear participated in the first class of Google Developers Launchpad Studio, focused on the applications of ML in healthcare and biotech.