The power of LLMs for differential diagnoses are remarkable, particularly when used in conjunction with a clinical decision support system. However, algorithms can also “hallucinate”, producing inaccurate or fabricated results which raises questions about how to address such instances.

Tackling these hallucinations requires the ability to tame algorithms and here validation becomes crucial, especially in professional domains where the stakes are high. It is likely that all current AI tools will eventually be superseded with models that meet our strict requirements and could place AI-assisted specialist HCPs in places where they are most needed at scale. Consequently, there is an inherent element of compassion in medical professions that needs to be considered.

Language is at the core of everything we do and plays a pivotal role in this context. All interfaces, whether genomic or mathematical, employ languages to transfer information. While AI can never (as far as we can deduce) replace the human touch in patient interactions, it can reduce time spent on administrative tasks and enhance communication language with more empathy during times of high stress. This could range from insurance denial to product leaflets, all subject to regulatory consent of course.

Medicine and healthcare have historically advanced by categorising everything, which forms the basis for all rules-based prescribing and reimbursement. However, we could move away from this rigid framework where in future, patient classification will be “fluid”, and AI tools' understanding of nuanced language could offer greater flexibility in decision-making.

To fulfil the promise of AI, underlying data requires a high degree of standardisation, but the models we have used so far have been primitive. Next generation models, powered by diverse data sources could unlock new ways of classifying language and develop new statistical methods. For instance, a conversation with a patient could be used to build a model specific to that patient, marking a fundamental shift in how interactions occur towards a more personalised experience.

Q&A

How likely is it that LLM outputs are trusted amongst specialists?

The tech sector’s exploration of medical-grade communication revealed that the highest bar of language lies with world-class professionals. It is also evident that domain-focused areas are already utilising AI tools. However, specialists often tend to verify the findings by using a search engine to find definitive answers. This suggests a degree of healthy scepticism in the reliability of LLM outputs.

The trust in LLM outputs could potentially be enhanced by training the models on data approved by a panel of experts. This method would ensure that the AI tool is learning from the most reliable sources, thereby increasing the confidence in its outputs among specialists.

Furthermore, tech companies are increasingly infiltrating vertical communities and assisting doctors in consuming information from their peers. This development could lead to a more widespread acceptance and trust in AI tools like LLMs, as they become further integrated into the daily workflow of medical professionals.

Will AI algorithms prefer medical evidence and form of rigid paradigms or will systems be able to shift and take into account divergent views?

AI algorithms' functionality is largely governed by the data they are trained on, and the design parameters established during their development. With prompt engineering, AI systems can provide outputs tailored to the style of the question asked, enabling a more user-specific response. As such, it is feasible to accommodate divergent views. Properly trained AI models should be capable of recognising the shifting nature of medical discoveries and adapt accordingly.

What do you see as real AI examples in pharma and healthcare?

Traditionally, pharma's approach has been push-oriented, where information is pushed out to the healthcare professionals (HCPs) by emails or a field force. However, with a growing focus on placing the customer at the centre, the industry is witnessing a shift in mindset towards a demand-led or pull-oriented strategy. AI plays a key role in this transition, analysing vast amounts of data to understand specific customer demands and enabling personalised content delivery based on individual needs and preferences.

How do we ensure policy is fit-for-purpose with the rapid progress of tech?

Ensuring policy is fit-for-purpose amidst rapid technological progress requires advancing real-time regulation and guideline creation. Life sciences has traditionally been driven by guideline-led science, which is slow and commonly inflexible. AI could play a pivotal role in policy discussions as its ability to perform live analytics and test models on the fly allows for immediate assessment of policy decisions. By leveraging AI's capabilities, we could create dynamic policies that keep pace with the swift advancements in technology, thus ensuring that they remain relevant and effective.

Any advice on how organisations can be trained and upskilled on adopting AI?

Organisations should first comprehend their capabilities and areas requiring integration. With AI interaction surfaces expanding due to enterprise development, understanding these interfaces is crucial. Some technical knowledge will always be required. Python is the primary language for interacting with AI models, but core skills in human intelligence, particularly logic, remain vital and the best skill to develop. Thus, a balanced focus on enhancing technical skills and nurturing human intelligence is key for effectively adopting AI.

AI is trained on historical data. How do you get a share-of-voice of new data if this info was not in the training data set?

AI is indeed trained on historical data, but it can still incorporate new data through continuous learning and adaptation. Pharmaceutical companies have very valuable data lakes that are continuously topped up. These can be integrated into AI models to enhance their value for patients. Moreover, AI will work to better understand the crux of patient queries and form a two-way data flow.

Furthermore, synthetic data can be generated and layered on top of historical data to augment datasets in areas with low data density. This enables AI to continually learn and adapt to new information, even if it was not included in the original training dataset.

Are there any territories where we cannot apply AI tools?

In theory, AI can be applied in any area where humans operate. The greatest challenge lies in developing the technology and data flows that feed effective AI models. The inflection point will be when AI starts creating new data models and statistics based on our needs and preferences. Even in areas with scarce data sources, strategies such as using analogues, mimics, and synthetic data can be employed to facilitate AI operations. Hence, while practical and ethical considerations may limit AI application in some areas, there are technically no absolute barriers to where AI tools can be applied.

Will AI do away with doctors?

While AI, particularly LLMs, are enhancing the speed and accuracy of diagnoses, doctors still play a pivotal role as decision-makers. The human touch, especially in delivering difficult diagnoses, is irreplaceable, ensuring a continued demand for human doctors. AI's increasing role in healthcare may, however, impact the length and nature of doctors' training.

If AI became the healthcare system rather than just assisting it, it would likely take over commoditised medicine, with the ability to communicate with inanimate objects (Internet of Things) to assist with medical care. This transition is already happening in small increments, but the timeline for full integration remains uncertain. Therefore, while AI will significantly transform healthcare delivery, it won't render doctors obsolete; instead, it will redefine their roles and daily activities.

NOTES

This article has been adapted from a panel discussion given by:

• Raja Shankar (Moderator), Managing Principal, EMEA Technology and IQVIA X
• Jas Dhalliwal, Google Global Principal Architect
• Prof. Thomas Szucs, Chairman of the board, Helsana & Co-director Genomic Medicine, Hirslanden Precise
• Deepanshu Mehta, Director, Data Science & Artificial Intelligence (DSAI) Novartis

at the Switzerland Strategy Conference held in Basel on 7 February 2024