As with any industry, healthcare and life sciences is a big market for technology – especially new technology such as machine learning and artificial intelligence. While the industry has pockets of opportunity for these new technologies, in this post, we take a look at some of the qualitative health tech trends and how they can provide better care for patients, be a better business for healthcare service providers and advance the industry as a whole.
Rising use cases in vaccine development
The global vaccine market was valued at about $35 billion even before the pandemic. Within months of the emergence of the Sars-Cov-2 virus on a global scale, vaccine trials had begun worldwide – a feat that would have surprised many just a few years ago. The fastest vaccine record is the mumps vaccine that took four years from sample collection to a marketed vaccine.
A large part of vaccine development involves analyzing messy, unformatted, and non-standard clinical trial data in various formats, research to find the right part of the virus that can be targeted and predict how future mutations will form a causality with vaccine efficacy.
The rapid development of AI and ML technologies, especially in the last five years, has largely contributed to a working vaccine being available within a year of the virus being discovered. While it cannot aid in the human-intensive part of trials, computational models can substantially optimize the vaccine’s chances of success.
Typically vaccine development begins with understanding mountains of data about the virus and the response it evokes in our immune system. The human immune system recognizes thousands of virus components. Hence, there are potentially thousands of permutations and combinations of how the vaccine can neutralize the virus. This task is a classic use case for machine learning and AI. AI algorithms can predict which parts of the virus the immune system is most likely to recognize using datasets of known pathogens.
AI and ML models can also aid in understanding how and which parts of the virus are likely to mutate. As viruses make copies of themselves, they often make minor ‘coding errors,’ creating new mutations. These mutations often get recorded in Germany’s Global Initiative on Sharing All Influenza Data (GISAID). AI and ML models can query these databases to see how the changes affect vaccine delivery and efficacy.
While current AI models are not reliable enough to predict vaccine efficacy, trials are still a reality. However, AI and ML are tools that help speed up the computing and information processing at scale tasks of vaccine development. With the rapid maturity of these tools to derive meaningful inferences, researchers worldwide hope to one day sequence vaccines for HIV and other diseases of whom important information exists without a viable vaccine candidate.
Data unity and ML ops
Whether in clinical applications or the administrative side of healthcare, AI and ML have applications that essentially underpin the need for data. Worldwide there are more than a few in-production AI applications. However, the data collected by the healthcare administrators differ drastically between service providers.
While AI models are already doing their bit, models developed and deployed in controlled environments may show inferences at par with a clinician. However, when the same model is paired with a different data set, it shows subpar results. On the other hand, the clinician can go outside the hospital bounds and still be just as effective.
For AI models to reach this level of maturity, the underlying data needs to be in a form that is uniform across environments and is standardized for models to use them. Data operations are already forming a critical part of AI’s success story in healthcare as we go forward.
One of the critical areas of concern with facial recognition technology is the model’s racial bias, causing inferences with grave concerns. There are chances of bias creeping in at every stage of the analytics process – data collection and curation, algorithm development, inference and decision making, closed-loop feedback, and model tuning. It is essential to understand that humans are involved in every step of the process and cause this bias to creep in.
As these technologies mature, technology practitioners are now taking a deeper look at all the things that can go wrong because of the model inferences and underlying data. AI practitioners recommended attaching metrics to these causes and then analyzing and testing slices of data to ensure there is no bias creep.
This analysis ensures that the model gets validated early on and remains free from any bias. But, at the same time, it also creates an audit trail. This trail is an essential reference should the model ever be questioned in the media, in a more significant forum, or a legal proceeding.
Moving from AI-ready to AI-enabled
The healthcare industry in the US is complex. All patient information and their medical data are documented and available securely. However, the system does not record all qualitative interactions between medical staff and the patients. This data is vital, and without it, only half the story is visible.
The unavailability of this data is a concern. AI and ML models built on partial data are limited in efficacy and ability to deliver actionable inferences. These models may work right for patients getting care from a specific service provider. However, outside this microcosm, its performance can deteriorate significantly.
On the other hand, clinicians such as physicians and radiologists can deliver the same level of performance irrespective of where they are since they rely not only on their training but also on experience to derive inferences.
There is also a significant gap between proofs-of-concept and production-grade and ready models. Yes, the tools and techniques for data structuring and standardization have improved significantly. Yet, the road to a production-ready model that can work anywhere is still some ways to go.
Better data transparency and validation
Across many projects where medical staff such as doctors and physicians work with technology teams, there is a certain skepticism regarding the inferences delivered by the AI model. This skepticism is primarily due to the medical staff not being aware of how this inference is derived. This situation is especially true when inferences are in contradiction of a prevailing best practice.
Hence it is vital to ensure that there is transparency in the underlying data. However, this is easier said than done. Healthcare in the US is a highly regulated industry with strict protections around data access and its end-use. While a specific set of people can access this data, they must follow specific protocols to de-identify the patient from their data before technologists can use it anywhere.
While the technology can do this, the data itself can vary in its formats and standards and lead to errors. These kinds of leaks in data processing carry significant legal and financial consequences for healthcare providers.
Hence one of the critical trends for healthtech as an industry would be to ensure data security and make it available for the common good.
AI in the administrative side of healthcare
The non-medical side of healthcare, such as scheduling appointments, facilities, and machines, accounting, staff payroll, supply chain management, are essential cogs to a well-oiled machine.
With digitization technologies now becoming mature, service provider operations generate large amounts of data. This data is helpful in intelligently automating substantial parts of operations. In addition, service providers can achieve tangible savings by using experience from professionals working in the field with data on how these services are being consumed.
The rapid proliferation of cloud computing technologies such as hybrid clouds ensures that non-critical workloads can run on public clouds saving substantial costs in on-premise processing.
There is substantial interest in healthtech, especially in how new technologies such as AI and ML drive tangible value in patient care and improve business KPIs and shareholder value. However, some significant gaps need to be plugged. These include data unification and access, change in mindset where data breaks down traditional hierarchies, transparency of the data and the model, and elimination of bias.
AI applications in isolation are currently in practice today, and these will replicate quickly across other providers as well. However, a concerted effort to advance the industry to deliver tangible goals using this technology will focus on healthtech investments in the coming years.