Artificial intelligence and the future of medicine:a multidimensional analysis

Philippe Moingeon

1French Academy of Pharmacy,Paris 75006,France.

This multidimensional analysis of the impact of artificial intelligence on the future of medicine aims to give some clues on foreseen categories of applications as well as their societal implications in terms of risks/benefits.Artificial intelligence encompasses technologies recapitulating four dimensions of human intelligence,i.e.sensing,thinking,acting and learning.Intelligent machines are converging with advancing biotechnologies to shape the future of medicine,in synergy with continuous progress in our understanding of system biology,brain physiology,biology of aging,computational sciences and decision-making theories.Data-driven predictive models of health-related problems can be generated to inform decisions and actions,allowing to enhance productivity in new drug development,increase the cost-effectiveness of fully integrated health care systems and empower patients and healthy individuals to better manage their disease or their health,respectively.Consequently,the future will likely take the form of a computational precision medicine continuously informed by data capture and modeling to propose preventive measures or therapies precisely tailored to characteristics of each individual.

Key words:Artificial intelligence,Biotechnologies,Computational medicine,Disease model,Drug development,Patient empowerment,Precision medicine

Background

The evolution of medicine in the last decades has been driven by progress in scientific knowledge and technological advances fostering each other.This trend will expand in the 21stcentury in the form of continuous and exponential rupture innovations along multiple dimensions.The latter are already embodied in terminologies such as post-genomic medicine,precision medicine,regenerative medicine,nanomedicine or digital health.Further scientific knowledge likely to impact the future of medicine encompasses system biology approaches to physiological and pathophysiological processes,brain and cognition physiology,biology of aging,computational sciences and advanced analytics,as well as modern theories on risk assessment and decision-making.Those converging advances combined with tremendous capital investment in the knowledge economy,as well as the strong albeit ambivalent appetite of postmodern humans for technological innovation provide a context for exponential and irreversible innovation never experienced before in human history.Specifically,a technological revolution linked to the synergy between biotechnologies,life sciences and artificial intelligence(AI) together with expanding scientific knowledge is becoming central in shaping the future of medicine and health in the 21stcentury.In this context,this review aims to describe the converging forces paving the ground for the technological future of medicine,review some prominent applications of this foreseen revolution and discuss as well some of its societal implications.

Converging technologies behind artificial intelligence

Technologies can be defined as the development of products,goods and services as practical applications of advances in scientific knowledge [1].Technologies themselves encompass a wealth of converging domains of expertise.The post-genomic era at the turn of the 3rdmillennium has fostered numerous emerging biotechnologies allowing to create,manipulate and transform living organisms.The future of medicine is now most particularly being shaped by a convergence between such biotechnologies with what is broadly defined as artificial intelligence.The latter classically refers to theories,activities and computational tools aiming to develop intelligent machines [2].The idea that machines could possibly reproduce some aspects of human intelligence was raised by Alan Turing as early as 1950 [3].As of today,the goal is not only to reproduce,but also to surpass human cognition by developing intelligent machines with a computing power allowing them to process massive amounts of data.

To this aim,AI incorporates technologies recapitulating four dimensions of human intelligence,including sensing,thinking,acting and learning (Figure1) [4].On a concrete basis,AI can take the form of an intelligent machine,for example a robot equipped with cameras and sensors providing data on a specific environment or problem (sensing),computing those data to build up a model in order to inform a decision(thinking),then engaging into a specific task (acting)and further,drawing additional information to adjust its future “behaviour” when facing a comparable situation(learning) (Figure1).

AI has gone through several winter periods since the 1950s,but the technology is now prone for inducing a revolution in many sectors including medicine,in light of major advances in computing power as well as the availability of massive amounts of data.It is estimated that quantities of data generated worldwide have grown with a compounded annual growth rate of +40%between 2012 and 2020 [5],in parallel with a similarly impressive raise in capital investment in AI [6].

AI in Medicine

Applications of emerging technologies and scientific knowledge foreseen to contribute to the future of medicine are presented in Table1 and summarized as follows.

Sensing technologies:

Sensing will benefit from information and computing technologies allowing to acquire (sensors),store (data lakes,cloud computing) and share (5 G,blockchain)massive amounts of data.Current innovation further includes computer vision,AI-based image analyses,speech recognition and natural language processing(e.g.in the form of intelligent vocal assistants).Applications include wearables,devices,and many other connected objects becoming available to generate or capture data on the patient and his/her environment.Sensing technologies also encompass high resolution imaging systems providing 3D representations of specific organs or tissue lesions,as well as advanced diagnostic technologies based on miniaturized,multiplex,automatized platforms performing highthroughput measurements of numerous biological parameters out of a small biological sample (e.g.a blood drop obtained by finger pricking).

Technologies to perform an extensive biological profiling of patients are now commonly applied for research purposes,and in the short future will be further applicable to routine monitoring [7,8].Whole genome sequencing is becoming more common and the exome(i.e.the coding part of the human genome) can be sequenced for less than 1000 US $.Technologies are available to characterize microbiota (the so called

Figure1 Components of human and machine intelligences.Four important dimensions of human intelligence are recapitulated by intelligent machines.The latter include the sensory perception of the human body (a consciousness of the environment which involves sensing and thinking),as well as acting appropriately in the context of those perceived conditions (combining thinking and acting).The learning dimension of intelligent machines is critical to constantly update models by integrating new data,thereby deriving a better probabilistic representation of a problem to inform decision-making and predict the future.

Table1 Technologies and scientific knowledge foreseen to shape the future of medicine

human second genome) and assess the diversity and composition of microbial flora (bacteria,viruses,fungi)present at various mucosal surfaces to decipher how they educate the immune system and influence susceptibility to diseases or responses to treatment [9].Molecular profiling data are also commonly generated by genome-wide analyses of susceptibilities to a growing number of diseases,single cell-RNA sequencing (evolving towards spatial analysis of gene expression in tissues),deep immunophenotyping with multilabel flow cytometry,and more generally,powerful multi-omics (i.e.proteomics,transcriptomics,metabolomics,epigenetics) technologies providing comprehensive information on biological processes.

Altogether,those analytical technologies provide huge amounts of data documenting individual patient characteristics in terms of their physiology,the pathophysiology of their disease as well as their lifestyle,diet,exposure to environmental pollutants,pollens and pathogens.Current challenges to their routine implementation include cost and accuracy of those methods,complexities inherent to data analysis and interpretation,and the need to cope with regulations regarding the use of sensitive health data.

Thinking technologies

Thinking comprises feature identification,problemsolving,risk analysis,decision-making,as well as predictions about the future.Thinking technologies in the form of computer analytics boils down to establishing models of problems (as diverse as either disease heterogeneity,evolution of an epidemic outbreak,interaction between a target and a drug,3-dimensional shape of a tumor…) by integrating big and multimodal data.Reaching the computing power to allow such advanced analytics has been made possible by progress in hardware with the development of processors with increasing performance in mathematical computation (evolving from central processing units to graphic and tensor processing units)in the last 20 years [10].

Progress has also been made in the architecture of computing units,which can be assembled in successive layers to process data and progressively extract patterns in order to obtain a desired output in terms of predictive capacity [11].Such artificial “neural networks” are built up as layers of computing units (“neurons”) to approximate the functioning of the human brain,in that“neurons” within a layer are connected to those in subsequent layers [12].Weights can then be allocated to each layer by back propagation,so that the network reaches the desired output in terms of prediction.This neural network architecture has proven to be key to allow machine learning,thereby opening a multiplicity of applications for AI [11].

Further progress in advanced analytics could come in the future from using supports distinct from silicium,such as DNA,or from quantum computing taking advantage of specific properties of matter in the quantum physics world.

Acting technologies

Informed decisions made from data-driven models can then guide proper actions.In the field of health,action technologies can take the form of robots performing any specific and repetitive task for which they have been programmed,e.g.robotic prescription dispensing systems capable of storing,filling,capping,administering medications securely and efficiently to patients.As it applies to medicine,action technologies also encompass means to design,develop,produce,administer better treatments.Specifically,computational modeling of the specificities of each patient will in the near future enhance considerably the current trend of precision medicine,allowing to better customize treatment modalities by taking into account the physiology,physiopathology,risk factors,lifestyle and relationship to the environment of every single patient.

Learning technologies

Learning makes the process of sensing-thinking-acting iterative by continuously updating probabilistic models of a problem to make them more accurate as new data become available.Machine learning involves training a system rather than programming it.To this aim,predictive algorithms are classically refined using separate training and testing data sets so that they can make meaningful inferences beyond what they had been programmed for on the basis of the training set [13].Deep learning,classically obtained by convolutional neural networks relying upon multiple layers for feature extraction,consists in feeding intelligent machines with massive input data from a wide range of sources so that they achieve better-than-human performance on a specific task (e.g.image classification or voice recognition) [11].

Future directions include reinforced learning,which alludes to the capacity of intelligent machines to learn by themselves:a perfect example consists in algorithms mastering intelligent games (such as chess or go) and even surpassing the best human champions after playing millions of parties against themselves [14].Quantum machine learning is another emerging area generating considerable interest,as it should capitalize on the speed and accuracy of quantum computing.

Technology-driven applications in medicine

The convergence of new scientific knowledge and technologies is creating infinite opportunities to transform medicine.Here below,foreseen applications will be emphasized in three specific areas,including (i)enhancement of drug development,(ii) empowerment of patients and healthy individuals to manage their health and (iii) support for better cost-effective health systems.

Enhancement of drug development

The joint use of biotechnologies and AI allows to create disease models that will inform all steps of drug development and help positioning therapies in welldefined patient subpopulations [15].Such models are generated following extensive molecular profiling of patients using multi-omics technologies to represent diseases as endotypes defined based upon underlying pathophysiological mechanisms,as opposed to clinical phenotypes.In this approach,biological data obtained in the blood and/or target organs are combined with clinical data to help stratifying patients in homogeneous subgroups to reflect disease heterogeneity.Using a system biology approach,patient subgroups are then further characterized in terms of molecular pathways which are dysregulated in comparison with healthy individuals [10].The latter information can then be used for identifying and pre-validatingin silicotargets of therapeutic interest.

Structural and biophysical information on those targets is computed to design,select or repurpose candidate molecules likely to interact with the target,and further to infer compound activity (based on free energy prediction) [10,16,17].AI and computational analyses help to optimize drug candidates by guiding chemical synthesis/ modification and predicting both microsomal stability,absorption,distribution,metabolism,excretion and toxicity.AI is also used during drug development to build up PK/PD models,predict clinical efficacy using quantitative system pharmacology and design innovative clinical studies comprising virtual placebo groups or virtual patients…[16-18].

Obviously,drug candidates are selected and evaluated by conducting in parallel empirical wet lab or clinical activities,but the end result of introducing in silico approaches is a much faster drug development process with significant derisking of the target prior to initiating time consuming and costly preclinical and clinical activities.As an estimate,AI and digital technologies could reduce the timelines for discovery and preclinical stages by up to 15-fold [19].

Empowerment of patients and healthy individuals to manage their health

There is a growing demand from patients to play a role in deciding together with health care providers the best options for their treatment [20].The common use of devices,wearables and health applications encourages patients to engage in managing their disease,by capturing and transferring their own data (related to diet,sleep,heart function,oxygenation,stress,exercise...) in real time,through a smartphone to health care providers.Applications for self-monitoring of patients with chronic diseases can be as diverse as an “artificial pancreas” presenting as a pump administering insulin to diabetics in link with their glucose levels,or“intelligent” pills detected by a sensor on the abdomen when ingested to reinforce patient compliance.

New technologies are also helping patients to develop a better knowledge of their disease,through the internet or social networks.Well-structured patient associations are very active at influencing regulatory authorities and drug companies by expressing unmet needs and even contributing in designing clinical studies [20].Expert patients in some countries are certified by universities on the basis of their extensive knowledge of their disease,allowing them to counsel other patients.The raising engagement of patients in their own medical decisions,but also in the design and implementation of healthcare services,has been occurring globally in the last few years.The development of metrics and tools to quantify patient engagement has only been recently initiated [21].

As stated above,precision medicine informed by patient-specific big data is currently moving from one drug-fits-all patients towards more customized treatment modalities,in the form of stratified or even personalized medicine.Importantly,it will also become more predictive.Increasing awareness by individuals of their risk factors,for example in the form of genetic susceptibility to diseases will come from routine application of genome sequencing,Single nucleotide polymorphism (SNPs) analysis and extensive molecular profiling.Healthy individuals with a growing concern of aging in good health will be encouraged to manage their health by adjusting their lifestyle habits and controlling their diet to prevent metabolic diseases and cancers.To an extreme,transhumanism might convince perfectly healthy people to use converging technologies to enhance their body performance and cognitive abilities [22].

Support for more cost-effective health systems

Technologies allowing to share data and information better and faster can bring significant improvement of productivity and help aligning resources in all sectors.As such,technologies will foster health industries 4.0 by integrating activities within the supply chain (e.g.storage,manufacturing,quality control) with those in commercial operations (sales forecasting,enhanced customer relationship management).In hospitals,such technologies will provide end-to-end digital solutions to manage optimally fluxes of patients,products and services,from procurement to invoicing.This will significantly facilitate the patient journey (in terms of flow registration,triage,diagnostic,testing,consultation,check out and discharge) with a better automation and integration of each of those steps and less patient wait [23,24].Full AI adoption by hospitals is expected to raise the productivity by 40 to 50%,with up to 90% of jobs in healthcare within 20 years relying upon significant digital skills [6,24].

Also,AI and related new technologies will make it possible to automatize numerous tasks currently performed by humans,both physical and cognitive,and facilitate complex decision-making [25].Technologies will thus also be used for the support or even replacement of health care providers (e.g.for image analysis) [23].Robots will handle patients,perform surgery,administer drugs,handle waste.Autogenerated reports will be transmitted in real time from electronic patient records,and telemedicine and robotized surgery will considerably expand [24].The overall end results will be better productivity and decreasing costs in health systems,with savings estimated to reach up to 300 billion $ per year in the US alone [6].

Current implications of future technological development

Societal support

Most post-modern societies have adopted a responsible benefit-risk approach in their support of technological innovation.Given the tremendous economical and societal value of the knowledge economy based on scientific innovation and emerging technologies,this trend should expand exponentially within the 21stcentury.As an example,public and private capital investment in AI-related activities alone has tripled every 3 years in the last decade,to reach hundreds of billions of dollars worldwide [6,19].

The support of individuals to technological innovation is more ambiguous.On one hand,smartphones and information communication technologies have been broadly adopted by billions of individuals in less than two decades.Millenials,that is the individuals born in the 80s and the 90s have grown up with digital technologies and expanding biotechnologies.Those belonging to generation Y are willing to use technologies to better control their life,shape their body and enhance their cognitive abilities.In the meantime,others -or sometimes the very same individuals - are critical about rationality,distrust scientific expertise and manifest vaccine hesitancy.This ambivalence towards science and technology has recently been reinforced by fear of a massive foreseen replacement of human beings by robots or the potential misuse of biotechnologies and artificial intelligence to the sole benefit of a few and the detriment of many [25-27].

Ethics and control of innovation

To address such societal concerns in the forthcoming decades,it is obviously important to regulate and control technological innovation,including those technologies shaping the future of medicine.Ideally,shared regulations should be adopted at an international level,in a risk-benefit balanced approach.Also,control of technological innovation by experts from the scientific community has been in the past quite effective,with as an example the Asilomar moratorium self-imposed by leading scientists in the 80s in the field of recombinant biotechnologies.In the last few years recommendations and guidelines have also been established by international panels of expert scientists for research on stem cell biology,manipulation of human embryos,studies on highly pathogenic infectious agents,applications of DNA editing technologies [27,28],and more recently on AI and robotics [26,29].

Two areas deserve specific scrutiny to ensure that technological innovation will be used to the benefit of mankind without creating any existential risk.The first one encompasses existing and future biotechnologies to create and transform living organisms,together with the emerging capacity to identify individuals with either genetic susceptibilities to diseases or propensity to express a phenotypic trait or behaviour [27].When applied to adults or embryos,those technologies could lead to various forms of eugenics and social engineering.In addition,monitoring the risks of inappropriate applications of artificial intelligence (such as diffusion of fake news,manipulations of opinions,population surveillance,cyberattacks…) is critical to ensure that intelligent machines of the future will not be detrimental to mankind [26].Specific risks to be considered in the health sector include a potential massive job loss (up to 50%) linked with AI and robotics deployment [6],a wrong doing of intelligent machines programmed by humans (leading to defective diagnosis or injuries to patients),as well as alteration in traditional relationship between physician and patients[23-26].To mitigate those risks,shared rules for governance and guiding principles for the development and use of emerging technologies need to be regularly updated and enforced at an international level [26,29].

The strategic importance of data

The health sector is specifically anticipated to benefit from AI and related technologies in that huge amounts of data are available,e.g.in the form of medical records in hospitals and private practices,or real-world data documenting efficacy and safety of thousands of drugs being used in hundreds of millions of patients [24].Biomedical research keeps generating every day massive amounts of data,and will continue to do so at an exponentially growing rate throughout the 21stcentury.The vast majority (up to 90%) of such health data is unstructured,taking the form of scientific literature,images,videos,audios,website content…[5].Hence the growing interest in learning intelligent machines capable to extract meaningful information from such massive unstructured data.Integrating those data to create and refine models to better inform healthrelated decisions will be a source of tremendous value.Data related to health are particularly sensitive and should be protected as such.Due to the economical value of such data,regulations on ownership,privacy and use have been introduced to this aim in the US(Health Information Technology for Economic and Clinical Health Act) and in Europe (General Data Protection Regulation).The quality and integrity of health data are critical and there is currently an interest in filtering on findable,accessible,interoperable and reusable (FAIR) data,or making it FAIR by curation,pre-processing and specific governance (with established rules for quality,capture,storage,sharing).

Conclusion:converging intelligences as a critical success factor

The British physicist Steve Hawking (University of Cambridge,UK) forecasted that success in creating an effective AI could be the biggest event in the history of our civilization… or the worst.To twist the balance between risk and benefit for mankind in the right direction,ensuring the convergence between human and machine intelligences appears critical [23].Machines can provide unlimited memory as well as fast and considerable computing power to conduct nonsupervised analyses from massive multimodal data.Humans are good at extracting features and providing transparency in models.Importantly,human intelligence is also key to the ethical development of future applications of AI [26,29].Intelligent machines have been recently able to defeat the best human champions in complex games such as chess and go.Interesting enough however,human champions have been able to learn new strategic patterns in those games and thus,made substantial progress by training with intelligent machines.Thus,convergence of the intelligences is possible,and it should arguably be considered as the most important success factor to ensure that technoscientific progress will benefit mankind during the 21stcentury,most particularly by contributing as a powerful and positive force to the future of medicine.