AI in Healthcare: two hurdles to tackle

With the (upcoming) paradigm-shift towards value-based healthcare, the importance of AI in healthcare will continue to increase. We notice two challenges for implementing AI in healthcare: data privacy and human-machine interaction.

In October 2019, we organized a round table with medical directors of Belgian hospitals about AI in healthcareAt this round table we outlined recent breakthroughs in AI technologies and their medical applications, identified and discussed the challenges for further progress in medical AI systems, and briefly touched upon the economic and legal implications of AI in healthcare. The outline was structured around a recent review paper on AI in healthcare1. Two topics we touched uponare still relevant today: data privacy and human-machine interaction.

Data Privacy

Legitimate privacy concerns are habitually raised when AI is mentioned in a healthcare context. Currently the main actors in healthcare AI development are research institutes, startups, or big technology companies (e.g. Apple, Amazon, Google). This is somewhat contradictory since it implies that data leaves the hands of the traditional trustworthy custodian (e.g. medical doctors and hospitals). To protect the privacy of the patient behind the data, the current approach is to anonymize the data before sharing. Anonymization is often achieved by de-identification (removing names) and sub-sampling (only releasing a smaller portion of the data). A paper published in the summer of 2019 made clear that this anonymization, using the aforementioned techniques, is reversible and that patients can still be identified2. Leaving the initiative to third parties and/or sharing data with external partners might thus prove troublesome (and hard/impossible to make GDPR compliant). 


“Blockchain technology in combination with AI offers great potential for collaboration in healthcare”


This does not mean that we should avoid the use of AI in healthcare, but it does urge to reconsider the way in which data is used/shared for AI in healthcare. New techniques, based on blockchain, offer great potential to safeguard privacy and to put the traditional data custodians firmly in the driver seat. A recent review paper indicated that the use of blockchain in healthcare (research) is increasing exponentially3. Inherent to the blockchain technology are characteristics such as decentralization, transparency, and anonymization. All of these can be used in the healthcare realm as has been proven by another recent study4.

In the latter the authors propose TCLearn, a methodology to implement trusted coalitions for blockchain distributed learning. This technique allows partners (members of a coalition; Trusted Coalition Learning) to train and share a model without sharing the data used to optimize this model. Translated to the field, this could for example enable hospitals, each with their own small/limited dataset, to collaborate in jointly training a model that is strong and accurate enough to be clinically relevant. And to do so without sharing sensitive data or without their data ever leaving the premises. The blockchain technology in combination with AI thus offers great potential for collaboration in healthcare. 

Human – machine interaction

AI is both a source of untapped opportunity and of existential risk. It is not easy to do right, and even harder to implement correctly5. One of the earlier efforts from AI that reached mainstream media was the victory of Deep Blue over Kasparov in the game of Chess. In the years following Kasparov’s defeat, Kasparov’s law was put forward, also known as the missing middle. Herein the importance of the process that encompasses the human-machine interaction is stressed.  


“The missing middle: Regardless of the strength of the human or the AI application/model, on their own they will always be beaten by a team with a process that enables the human to augment his/her capabilities by the machine’s capabilities.” 


Recent evidence from AI applications in industry have shown that success is more likely when sufficient attention is placed on the integration of AI in the current processesand on the impact of the people collaborating with the developed AI applicationThe production of AI should therefore always be aligned with the consumption of AI, and care must be placed to ensure that AI solutions are adopted effectively and pervasively. Or as the CIO of Roche Diagnostics has put it “AI is not in itself, a separate agenda. It is a subset of the tooling and the capabilities and methods we are using to pursue strategic objectives.”  Several studies support this hypothesis and it is one we as Möbius underline. AI will never reach its full potential unless integrated in the workflow, aligned with the work floor and with attention for the customer experience.  

AI requires a holistic, multi-dimensional approach. We strive towards a strong coupling between human and machine, in the form of Augmented Intelligence (AI). Besides extensive in-house AI knowledge, Möbius has a long track record in facilitating collaboration, enhancing customer experience, process optimization and change management business consulting. 

We are enthusiastic about the potential of AI to offer novel solutions for deep-rooted challenges in the healthcare realm. We encourage your feedback and partnership, to together build, adapt and share our collective path forward. 

 

References: 

  1. K-H Yu et al. 2018, Artificial Intelligence in healthcare, Nature Biomedical Engineering. 
  2. L Rocher et al. 2019, Estimating the success of re-identifications in incomplete datasets using generative models, Nature communications. 
  3. A Hasselgren et al. 2019, Blockchain in healthcare and health sciences – A scoping review, Int. J. of Medical Informatics. 
  4. S Lugan et al. 2019, Secure Architectures implementing trusted coalitions for blockchained distributed learning (TCLearn), IEEE Access. 
  5. C Longoni et al. 2019, Resistance to medical artificial intelligence, JCR. 

Thanks for reading

Share blog