At a Glance
- AI can improve disease diagnosis, treatment plans, and regulatory compliance through advanced data pattern recognition.
- Generative AI models are broadening the scope of AI applications in medical devices.
- Despite its vast potential, the integration of AI in healthcare raises significant challenges and ethical considerations.
Given the vast amounts of data available, including raw measurements, diagnostic information, treatment plans, and regulatory guidelines, the biomedical technologies sector stands to gain immensely from artificial intelligence (AI), particularly machine learning (ML).
ML, at its core, learns from training datasets to identify patterns, which can then be applied to new input data to make direct inferences. For instance, if specific body scans frequently result in a particular diagnosis, ML can be used to quickly provide that diagnosis when similar scans are encountered, thus aiding in disease diagnosis.
Similarly, if certain symptoms often lead to a specific drug prescription, ML can be used to promptly suggest that prescription when those symptoms appear, thereby assisting in disease treatment. Additionally, if healthcare records missing certain values typically require updates to meet regulatory standards, ML can be used to swiftly recommend those updates when similar records are found, thus supporting regulatory compliance.
How GenAI is opening doors in medtech
Over the years, ML techniques have indeed been applied to various types of data in biomedical technologies to draw direct inferences. However, in the past year, more powerful ML models with huge sets of parameters have been developed from massive training datasets.
These models can detect numerous subtle patterns in the data and generate expressions that mimic human language based on these patterns and new input data. These advanced models are known as generative AI (genAI) models.
While the fundamental inference approach remains unchanged, the ability to generate humanlike expressions has expanded the scope of service of these ML models. For example, a straightforward application of the conversational capability is to clinical treatment devices, such as CPAP machines, dialysis machines, and infusion pumps, to improve the human-device interface and further personalize the user experience.
More broadly, given the detected patterns, new input data can be summarized by eliminating portions that do not match the detected patterns, and are thus considered less informative. This allows ML to assist in digesting research papers, medical histories, government regulations, insurance claims, and other lengthy or complex documents.
In another example, if a detected pattern indicates that feature A has a specific relationship with feature B, new input data inquiring about this relationship with feature A can be addressed with information related to feature B. This enables ML to answer questions regarding various biomedical processes that are well documented.
Additionally, if patterns show that feature A is related to features B and C in similar ways, new input data seeking an alternative to feature B for feature A can be addressed with information related to feature C. This allows ML to assist in formulating alternative diagnoses, trying substitute drugs, designing replacement medical device components, or providing other innovative solutions.
Taking advantage of specialized languages
To date, these genAI models have primarily been applied to input data that humans can easily comprehend, such as natural language text and images representing familiar concepts like colors, shapes, and objects. However, a promising direction is to extend the application of genAI models to complex, high-dimensional data, which are often domain-specific and not immediately interpretable.
It is important to note that domain-specific data is not just natural language text and images related to a particular field; it can include arbitrary arrangements of specific letters, numbers, or symbols in specialized languages. Example data structures for such datasets include the Protein Data Bank (PDB) format used to describe three-dimensional structures of molecules and the STEP.
ISO 10303 (STEP) format used to represent three-dimensional physical objects, including components of a medical device. Such domain-specific data offers advantages over natural language text and images because it contains less noise unrelated to the domain of interest but instead includes more built-in emphasis on relevant features and relationships.
Such domain-specific data might have already been explored using traditional ML models. For instance, one-dimensional protein sequences have been analyzed and mapped to three-dimensional protein structures, which are often linked to protein functions.
Clinical trial data have also been integrated into medical device simulations for early detection, and hopefully prevention, of adverse events, such as catastrophic seizures and stroke. By venturing into these intricate, domain-specific data types, genAI can unlock new possibilities and insights.
First, genAI models can now detect finer and subtler patterns in complex data. For example, these capabilities enable the development of more accurate mappings from protein sequences to protein structures, or from protein structures to their functions, especially when experimentally determined structural data is available.
These capabilities also permit more detailed transformation from component sensor data to component interactions or from component interactions to system-wide reactions.
Nvidia's Clara platform is a notable example of such efforts. Furthermore, these capabilities can allow for the creation of mappings from protein sequences directly to protein functions or facilitate the projections from sensor data directly to improvements in medical devices, each of which is understandably a less direct and more convoluted process.
Additionally, genAI models can synthesize new complex data based on the detected patterns. For instance, once the portions of a protein structure corresponding to important protein functions are identified, genAI models can propose a "summary" protein structure, which is a more compact version of the original. When this protein structure corresponds to an existing drug, the more compact version might lead to a new drug with reduced interactions and fewer side effects.
In the context of device configuration, the more compact version might similarly lead to a new medical device that is more lightweight and easier to apply to the human body.
Furthermore, genAI models can propose a "substitute" protein structure or device configuration by replacing one original component with a novel one that satisfies all intercomponent patterns, ensuring the original structure or configuration is not disrupted.
This substitute can not only correspond to a similar protein function or device utility as the replaced component but also provides additional benefits. Therefore, instead of merely suggesting an existing alternative, (in natural language), genAI models can potentially construct a solid new alternative.
A few cautionary thoughts on advancing AI technology
With tremendous promise comes incredible challenge. Among the primary concerns with advancing AI technology are explainability and transparency, as emphasized in major AI regulations and guidelines such as the White House's Blueprint for an AI Bill of Rights and the European Union's Artificial Intelligence Act.
As ML models grow increasingly complex, understanding their algorithmic workings turns more difficult. The growing complexity of the data flowing through these ML models further exacerbates the issue.
It is true that different types of output data from ML models raise varying levels of concern. For instance, when the output data pertains to scientific constructs like protein docking mechanisms or binding affinities, the experimentally verifiable nature makes it less critical to understand how the data was derived. However, when the output data involves medical recommendations, the validity or effectiveness of the data is either challenging to verify or needs to be verified as soon as possible. In such ases, understanding the derivation of the output data becomes crucial.
Moreover, processes and operations concerning personal health have long been subject to stringent regulations. The same fundamental concerns, particularly privacy and liability, become even more pronounced with the advancement of AI technology.
Privacy faces escalating challenges as the chain of custody for personal data gets increasingly complex, involving various parties including third-party AI developers who specialize in handling vast amounts of training data and machine learning models. This clearly results in more opportunities for privacy invasion and security breach.
Additionally, issues of ownership and liability are becoming increasingly ambiguous as more powerful AI technology leads to greater automation with minimal human oversight.
Humans will persist in enhancing computer capabilities to simplify and improve our lives. With the continuous progress of AI, we are embarking on an exhilarating journey that promises to revolutionize medical devices, life sciences, and numerous other industries.
Originally published by MD+DI.
Disclaimer: This Alert has been prepared and published for informational purposes only and is not offered, nor should be construed, as legal advice. For more information, please see the firm's full disclaimer.