Patient Centric Healthcare

    The Role of AI + Data Science



    Download PDF VersionClose



    Download the Whitepaper

    The shift to patient centricity

    Today’s consumers expect choice and personalisation, creating huge disruptive pressure on the healthcare sector. These changes – enabled by a superconvergence of companies, technology, data, and connectivity – are affecting all aspects of global business, and healthcare is no exception.

    Netflix and Spotify disrupted complacent industries by putting consumer needs and convenience at their centre, using clever algorithms to make mass market services more personal. The life sciences and healthcare industries are undergoing a similar shift, albeit with more complex challenges.

    But at the heart of this shift is the same end user demand that is disrupting all industries – customer centricity, which in healthcare means patient centricity.


    Profits from Blockbuster Drugs will be hit first

    A patient centric experience is a complex, multi-faceted one. It’s about personalising drugs and treatment regimens for specific groups, adjusted for genetic makeup, lifestyle, and medical history. It’s about faster and more accurate diagnoses, enabling more tailored responses. It means a more personalised approach to remote care and seeking to prevent illness through health and lifestyle support.

    At the heart of patient centric revolution is the effective use of research and patient data, applying Artificial Intelligence (AI) and data science to best harness it. Despite the enthusiasm for AI, this is hard to do well and there is a high risk of failure.

    AI can analyse huge amounts of complex data and spot meaningful connections that humans would struggle to see (e.g. between human activities or drug chemistry, and outcomes). This is increasing the capability to decouple complex variables and reach a far more nuanced understanding of cause and effect. Using AI allows companies to better understand what their data is really telling them and how it’s affected by different situations, rather than relying on generalisations.

    There’s a huge opportunity to develop drug variants suited to different groups and lifestyles

    For pharma and biotech, this will mean becoming smarter with data, developing therapeutics which meet the needs of individuals rather than populations. Whilst the future may have fewer blockbuster drugs, there’s a huge opportunity to develop drug/regimen variants suited to different groups and lifestyles as we better understand how health, lifestyle, and other variables affect efficacy.

    Diagnostics companies will harness new data sources to develop decision and diagnostic support for doctors and consumers. Meanwhile, tech companies are racing to use AI to develop tools that deliver personalised advice to help people lead healthier happier lives or maximise their treatment regimens (e.g. Digital Therapeutics – provision of an app with a pill). These developments are attracting considerable interest from health services and insurance companies, as well as end users.

    Naturally, people want to avoid being ill and minimise the impact of ill health. They – and the health services that support them – spend a lot of money on both. But they also have more choice than ever before. The industry needs to keep innovating at the top of their game to stay competitive. This will be challenging, but hugely lucrative for those prepared to take advantage.

    This article will explore the opportunities to harness AI and data science to deliver patient centric innovations across the value chain, and explore the practicalities of building and deploying AI in this context.

    What is driving patient centricity?

    The health sector is made up of a wide range of organisations. Pharma, biotech, and medical device and diagnostics companies exist at one end, and health services, private hospitals and insurance companies at the other. Recently added to this mix are a huge number of tech companies offering new solutions to both.

    Before we explore how AI and data science can benefit these different industry verticals, let’s briefly look at the global trends driving change across healthcare. Remember, these trends are interwoven and represent a superconvergence that cannot be viewed in isolation.


    Consumer choice

    Consumers have an increasing amount of choice. They have on-demand access to films, TV, music, banking, and mobility, it’s no wonder they expect more choice in healthcare.

    Consumers have access to unlimited information (not all of high quality) to help them understand their own bodies, and they’re using it – 64% of patients seek information about conditions outside of physician visits. Apps are offering personalised health plans. Home testing kits offer rapid diagnoses, which some see as complementary, or an alternative, to doctor diagnosis. Future patients are likely to be more questioning, have higher expectations, but also be more open to new approaches.


    Preventative health

    Naturally, patients prefer not to be ill and to live more comfortably when they are. At the same time, evaluating, treat, and maintaining patients in hospitals is expensive.

    There is therefore a growing drive to reduce people’s need for care. In fact, an estimated 50% of healthcare services are due to move from hospitals and clinics to homes and communities over the next decade.

    As the connected world opens channels for remote care, evidence-based approaches that keep people from becoming ill or make conditions more manageable are likely to see increased uptake.

    Precision medicine

    There’s growing interest in treating subgroups rather than taking a one-size-fits-all approach.

    With more data and better tools to explore it, we can more accurately predict and measure how variations in drug chemistry or treatment regimens affect different groups depending on their genetics, age, or lifestyle. This will allow companies to capture sections of an existing market by providing them with more effective treatments.


    Wearables, sensors, and other mobile technologies capture health data in ways previously impossible. This allows unprecedented insight into the health of individuals, and unprecedented ability to develop targeted new products and personalised healthcare plans.

    The associated connectivity of these devices opens new possibilities to monitor effectiveness and side effects, and nudge people towards good behaviours, from healthy lifestyles to drug adherence.

    Storage and processing power

    Machine learning, a subset of AI, has long been used in R&D. But advances in algorithms and new, more sophisticated tools are adding ability to tackle more complex and ambitious tasks. These advances have been made possible thanks to the explosion in the amount of data available to process, the ability to store it, and the exponential increases in compute power.

    Looking over these trends, it’s clear that companies will need to move beyond the traditional healthcare industry model. They’ll need to integrate their capabilities and technologies, innovating together in order to meet patient demand and addressing the disruption required.

    Companies who have traditionally created products to sell to healthcare providers, must now consider the end user too. They must be able to extend their business models from B2B to B2B2C. This will be a new mindset for many.

    AI in healthcare: The Wild West

    AI classifies data based on the relationships between many different interconnected factors. Unlike traditional software, which follow rules defined by software engineers, AI automatically formulates the rules from the data it’s trained with.

    For example, an AI model fed large numbers of images of different skin rashes can learn to spot each type based on their unique combination of characteristics without being told what a particular rash looks like.

    It can also find new links: a model can be told what pharma research is trying to achieve, then analyse molecule libraries to identify likely candidates without being explicitly trained on what to look for. In some cases, this can lead to approaches that no human would identify.

    NASA used generative algorithms to design an antenna against a set of criteria. The result would never have occurred to a human, but was better than anything else they came up with. Similar approaches are being used in drug design.

    In this complex landscape, it’s hard to cut through the noise and understand what AI can really do

    AI is also good at isolating complex variables. For example, an AI can model the implications of multiple-drug regimens. For humans, when looking at patients taking multiple drugs, it’s too hard to isolate all factors and conclude that a particular interaction was having an adverse effect. But this is where deep learning shines. With enough data from large populations, AI can spot weak signals that show how and when specific combinations of factors lead to specific outcomes.

    Right now, AI is like The Wild West. There are lots of promised solutions, but little clarity. In some cases, AI is already delivering value, usually to projects where AI has been purpose built for that challenge. There are many promising transformational applications of AI, though some may not be possible for years. Then there are other pursuits that are no more than overblown marketing claims.

    In this complex landscape, it’s hard to cut through the noise and understand what AI can really do for an organisation. AI has huge potential, but – for now at least – it’s rarely easy to implement and only works with the right data, models, training, and deployment.

    The cross-cutting impact of digitalisation, AI, and data science

    Patient centricity is driving change across all areas of healthcare. Ensuring AI and data science can deliver the outcomes expected will be key at a programme and project level.

    In this section, we’ll look at real examples of where AI and data science are delivering impact in three major areas – including several in which Tessella or its parent company, Altran, were involved – and assess how we can help AI move from a technology trigger to a real force for productivity.

    How do we seize the opportunity?



    Case studies

    Disruptive discovery and clinical innovation

    AI can look for hidden patterns in complex sets of diverse data – from chemical analysis, to past clinical trial data, to user feedback – and spot insights that will guide more focussed and targeted drug development.

    Detailed patient data from clinical trials

    Many clinical trials of ongoing drug regimens require regular check-ups, but patients often behave differently under observation, skewing results. For a new arthritis drug, a major pharma company is using medical grade sensors to assess people in their home, to measure factors such as speed of movement and transitions, which they can then link to drug performance. By applying AI, it’s possible to establish baseline patterns of behaviours and spot subtle improvements.

    Decoupling effects of stress from drug research

    Verum, designed by Cambridge Consultants, is a wearable which remotely monitors stress levels. It monitors voice and electromyography (EMG) and applies machine learning to predict stress levels. Since stress can have an impact on drug efficacy, this allows researchers to understand the impact of stress during trials, which in turn allows it to help support patients and improve trial design to decouple the effects of stress from its results.

    Peace of mind about skin cancer

    SkinVision is an app that checks for signs of skin cancer using your phone camera with instant results. Trained on a large database of skin cancer images, its AI can spot suspicious signs which should then be checked by a specialist. Such image analysis relies on very subtle physiological correlations, which would not be possible without AI.


    Personalised, evidence-based diagnostics

    AI can look for hidden patterns in complex sets of diverse data – from chemical analysis, to past clinical trial data, to user feedback – and spot insights that will guide more focussed and targeted drug development.

    Precise rapid diagnosis of infectious disease

    Stat-Dx checks for every possible pathogen of a specific syndrome in an hour, allowing a rapid diagnosis. Usually such tests involve sending samples for lab analysis, meaning effective treatments are delayed or incorrect antibiotics are prescribed.

    Stat-Dx’s platform, which has since achieved regulatory approval, applies a chemical process to the sample which amplify pathogens causing them to fluoresce. Sophisticated algorithm and software techniques were developed, with support from Tessella, to link the fluorescence pattern to the presence of a specific pathogen, achieving diagnosis accuracy of more than 98%. The system gives doctors a rapid diagnosis of the exact infection, allowing them to prescribe the right treatment immediately.

    Diagnosing respiratory disease by the sound of your cough

    ResApp uses a microphone to listen to coughs and identify different types of respiratory disease. A machine learning app trained on a large dataset of coughs linked to different diseases can distinguish between different types. It’s currently used by clinicians to support diagnoses but may be rolled out to consumers in future

    Diagnosing gastrointestinal disease by the smell of your breath

    SniffPhone is investigating using a small plug-in module for a smart phone that can detect disease from exhaled breath. Breath samples are digitised and key parameters are compared to the app’s database. It then detects subtle indicators of a wide range of gastrointestinal diseases, including various cancers.


    Smarter healthcare and wellbeing

    New sensor and communications technology (e.g. smart watches) allows health data to be collected from people as they go about their lives. AI can learn from this data at a granular level to understand how healthcare approaches affect different individuals. This can then be used to make recommendations to optimise healthcare regimes or promote healthy living and disease prevention.

    Improving quality of life for diabetics

    M4P, a French project led by Altran, is establishing the ‘Diatabase’, which is capturing information on the life and health of diabetes sufferers. As the database grows, it’ll establish how a patient’s environment affects disease progression, treatment effectiveness, and impact of other diseases.

    With added analytics and interfaces, this will let doctors identify optimal, personalised treatment and lifestyle regimens. Once proven with diabetes, the same architecture can also be used for other diseases.

    Monitoring blood pressure from your smartphone

    Riva Digital uses your smartphone camera to measure the colour of blood flowing through your fingertip and thus deduce blood pressure. Using AI, it can sense very slight variations in colour and link them to pressure.

    The first version of the app warns of danger signs, but future interrelations will augment this with healthy living advice. Further AI could be applied to monitor diabetes (via changes in blood flow).

    Personalised nutrition and weight loss based on your metabolism

    Lumen is a metabolism measurement device, integrated with coaching algorithms. It measures your breath to understand individual metabolism based on the gas composition. It then delivers personalised weight loss and fitness advice based on what your body needs at that moment, rather than one-size-fits-all plans.


    Smarter healthcare outcomes: two use cases for delivering AI in practice

    Some organisations are already delivering successful AI projects. Here we look at two projects Tessella supported, to explore how the company approached planning, data gathering, model design and project delivery to produce tangible outcomes that had a demonstrable business benefit.

    Use case 1: AI-enabled drug design

    A major pharma company is investigating how AI can improve its drug design process.

    A common drug design approach is to use a higher-level description of what they want a molecule to look like. One such description is the reduced graph, which involves specifying in chemistry terminology what structure the molecule should have. For example: “an aromatic ring connected to a linker, which in turn is connected to an aliphatic ring acceptor, which in turn will potentially be connected to several other molecular substructures with different characterisations.”

    This high-level description limits the search for molecules to those which meet specified criteria, i.e. having a similar structure to a known active compound. Creating a reduced graph for a known molecule is easy; the bigger challenge is the opposite process – finding good potential molecules which match the desired reduced graph. It’s a bit like buying a house. If your criterion is “any house”, you’ll never find what you’re looking for. But if you specify location, number of bedrooms, and price, you have a better chance.

    As drug research becomes more personalised, the ability to reduce time finding the right molecules will become increasingly valuable.

    Specifying the reduced graph of a molecule is like providing a detailed layout for your ideal home. However, while there are a million or so property ads online, the number of molecules available for drug design is around 1,060 with the overwhelming majority never having been synthesised.

    This challenge of generating a set of candidate molecules from a reduced graph description is something AI can assist. Remarkably, it was discovered that this problem can aid a completely separate AI challenge: translating languages.

    Language translation takes advantage of two cutting-edge developments in neural networks: “sequence-to-sequence learning” and “attention mechanisms”.

    Sequence-to-sequence learning takes a sequence of words – a sentence in English – and outputs another sequence of words – a translation in French. Languages have very different structures, which is why successful machine learning approaches consider sentences in their entirety and generate a new sentence which captures the whole meaning.

    Attention mechanisms allow models to focus on particular words in the input sentence.

    It’s, of course, also useful to know that particular words in each language relate to each other, and this is where the attention mechanism comes in. Attention mechanisms allow the model to focus on particular words in the input sentence when generating particular words in the output.

    Together, this allows translations in which the right words are selected, but also capture the correct overall meaning.

    A molecule can be represented as a text sequence using a SMILES string. The same is true of the high-level reduced graph capturing the outline of what the molecule should look like. It’s possible to apply the same basic principles of language translation to “translate” the outline of a molecule into a specified novel molecule matching the outline. In other words, to predict a molecule to match requirements.

    All you need is a dataset with hundreds of thousands of molecules and their equivalent reduced graph outline to train the AI system. Fortunately, there are huge datasets of molecules are readily available and generating high-level descriptions of a complete molecule is relatively easy. For any given reduced graph, the AI can propose new molecules which match the specification, which chemists can use to guide their search for the next drug candidate.

    To validate the model, you feed in high-level reduced graphs of drug candidates from published literature – ie with existing answers – that the system has not seen. If the AI could take these high-level descriptions and generate a known active compound, this is a great indication of its value in future discovery programmes. In work published in the Journal of Chemical Information and Modelling, this test was performed with 20 different known active molecules. In most cases, a known active compound was generated, and in all others, molecules generated were similar to a known active compound.

    Such approaches can benefit all aspects of drug design. As drug research becomes more personalised, the ability to reduce time finding the right molecules will become increasingly valuable.


    Use case 2: diagnosing rare diseases with AI

    Anna Litvak-Hinenzon and DS team (Tessella), John Reynders and DGB team (Alexion)

    Quick and Accurate Diagnosis saves lives

    At six days old, Maverick Coltrin developed seizures. Doctors were baffled, and his parents feared for his life, and conventional diagnostics had failed to reach a conclusion. The medical team at the Rady Children’s hospital suggested Maverick take part in an innovative digital diagnostics study. The approach used data gathered from the latest genetic sequencing techniques, then applied Artificial Intelligence (AI) to match it to existing research and diagnostics information sources about rare genetic diseases.

    The results came back, showing that Maverick had a rare form of epilepsy, but one that was easily treatable with vitamin B6 supplements. His life was transformed. The technique used to diagnose Maverick is part of a growing field of digital diagnostics made possible by the combination of sophisticated AI and the ever increasing availability of genetic and patient data.

    Why the time is right for AI-assisted diagnostics

    These new approaches to diagnostics are possible because of the ever-increasing availability of data and sophisticated low-cost AI tools. Every year the cost and time to produce genetic sequencing data are more economical and faster than the previous year. Patient electronic medical records (EMR), digital scans, and digitized research outcomes are now more readily available than ever before. The most advanced AI tools ever developed are widely accessible in an open-source format and often pre-installed on powerful yet reasonably priced cloud platforms.

    Every year the cost and time to produce genetic sequencing data are more economical and faster than the previous year

    The secret sauce

    It is essential to understand that developing AI diagnostics systems is not as straight forward as just combining cutting edge AI technology with vast quantities of data. Generating insight from clinical and healthcare data is challenging work; it needs medical science expertise, world-class data science technology, and talent, as well as the IT experience to integrate emerging and novel technologies into robust operational systems.

    The vast history of medical literature and genetic data does not present itself in an easy to navigate spreadsheet – but appears in many different and complex formats, from research papers to handwritten notes, to images. Comparing and extracting this kind of data demands a highly specific approach to the design of AI and machine learning algorithms. This kind of work is far removed from the world of e-commerce or consumer AI, where simple correlations can deliver a “good enough” return. The need for the highest accuracy and a rigorous proof of causation from complex data sets requires a very different breed and blend of data science.

    In this field, much of the work utilizes Deep Learning, a form of AI whereby the model learns to interpret relationships in large and complex sets of data. Presented with enough data, it gradually learns to identify what is essential and how information relates to each other. It typically involves training a multi-layer artificial neural network on a labelled dataset. With enough training data, a well-designed deep learning model will classify information as accurately as human experts, in a fraction of the time. Other advanced AI and machine learning techniques may combine several models together to create sophisticated learning capabilities that, when applied to a variety of data sources, produce highly accurate predictions.

    Who is doing it well?

    Many organizations are doing this well. For example, a system developed in a collaboration between the Manton Center for Orphan Disease Research, part of the Boston Children’s Hospital, and Alexion is able to produce a highly personalized sequence of clinical questions to aid diagnosis of rare-diseases. This system combines the diagnostic knowledge of healthcare professionals, a global disease database, and the specifics of a patient’s personal medical history to go through a series of questions that step by step narrows down the highest likelihood indication.

    The resulting AI platform, developed by Alexion with their affiliates, uses various AI algorithms including the “20 questions game” format. This platform has the potential to place the global knowledge of rare disease into the hands of every clinician. They then combine this with their personal experience and expertise in patient interaction to significantly increase the probability of diagnostic success. Rare diseases present a specific challenge to AI systems – AI systems need large quantities of data to ‘learn,’ but by definition, large data sets are not available for rare diseases.

    This platform has the potential to place the global knowledge of rare disease into the hands of every clinician

    Another Alexion affiliate, The Rady Children’s Institute, has gone even further, leading it to receive a Guinness World Record for a rapid under 24 hours diagnosis, using genome sequencing as an input to the AI diagnosis. An AI engine analysed the digitized results and identified the overlap between the specific observations of a given child’s illness with a reference set of expected observations covering all genetic diseases. The key to the automated diagnosis is combining genome sequencing, phenotypes and Natural Language Processing. The patient’s electronic medical records (EMR) are processed to identify and extract observed disease features; these are then compared with the results of genetic sequencing. In this race against time, the manual review of health records takes precious hours from an overloaded medical professional and is prone to personal subjectivity. AI diagnosis support, which combines the evidence in the gene sequence and health records, not only reduces the elapsed time, it can combat this bias.

    Companies like alexion, working with their affiliates, are looking to bring ai-assisted diagnostics to an ever-wider range of diseases 26 27 get the planning focus right

    Future AI diagnostic

    Companies like Alexion, working with their affiliates, are looking to bring AI-assisted diagnostics to an ever-wider range of diseases. They will look to make greater use of phenotypes and genetics data to reach this goal.

    The challenge for the Alexion data sciences team and those like Tessella working with them is to master and integrate an extensive range of highly specialized algorithmic techniques, known under the generic term of AI. This challenge has been met through an active collaboration between internal and external data scientists working closely with physicians and healthcare experts. New specialists are able to join the core team rapidly and communicate clearly how complex technical choices will play out to the medical, business, data, and IT subject matter experts.


    Thanks to AI diagnostics, new levels of speed and accuracy are possible, helping doctors make correct diagnoses and expanding the likelihood that they will spot more unusual conditions. The upshot will be faster interventions and increased accurate diagnoses. This will reduce costs of readmissions and increase the chance of initial treatment being successful; most importantly, it will save lives.

    How to deliver AI and data science projects

    As you can see, there is huge potential for patient centric healthcare. Both AI and data science play a critical role in delivering its prospects.

    Exploring the potential of AI in an organisation comes with significant challenges, as does industrialising it. Below, we provide a guide to building and implementing AI to support the journey to patient centricity.

    Decisions first, data last

    Start by defining what you’re trying to achieve. Successful AI programmes start by identifying decisions needed to solve problems. They should ask how those decisions – whether they’re human or automated decisions – could benefit from greater insight and what these insights might be.

    Only then start identifying what data you need to achieve this. Sticking sensors in and seeing what insights data throws up may yield some results, but it’s unlikely to be cost-effective and probably won’t give you the results you want. Instead, plan to capture the data you need to drive decisions.

    Stay agile

    In the uncharted landscape of AI, heavy planning, long development cycles, and huge investments all increase risks. Successful AI pioneers have a clear objective, but explore different approaches through agile, iterative, real-world experimentation.

    Take a connected approach

    Ensure everyone understands what they’re trying to achieve, not just what they’re doing. In large, complex organisations, it’s common to assign actions without true understanding. For example, a department will be told to setup sensing devices to capture data without fully understanding how that data will be used. Years later, the data team will come to that data and find it isn’t actually what they needed.

    Employ the right expertise

    AI is complex. It needs people, not just technology. AI is a new technology, and companies often don’t have the specialist skills to harness its potential.

    Successful AI projects need knowledge of relevant AI techniques (of which there are many) and of devices and data. They also need domain expertise to bring understanding of what the data means in the real world, e.g. biochemical interactions or human biological functions. And, in the patient centric world, they need experts who understand what this means for patient quality of life.

    This path from data collection to meaningful human outcomes must all come together, which means also employing translators who can bridge the language gap between different experts – physicians, chemists, businesses, doctors, carers, and patients.


    Capture the right data

    Health data is complex, varied, and often poorly structured. It’s a mix which includes data from chemistry, biology, genetics, human activity, medical images, medical notes, and electronic health records. To get good results, you need to feed good data into the model.

    This is often challenging to capture reliably. Handwritten doctor’s notes can be hard to accurately digitise. Data from consumer smartphones introduce risks from uncontrolled variables, firmware updates, raw data, and poor-quality data. This can still be useful – top end smart watches, for example, are great at data collection – but it’s important to setup data collection correctly to deliver the type and quality of data you want, and to understand its limitations.

    Validate your data

    In some cases, it’s necessary to verify that your data is actually a ‘real effect’. For example, a wearable sensor may infer that a data pattern means the patient is stressed. But to be certain, you need to setup trials where you know the patient is stressed and ensure your measurements consistently correlate with other reliable stress indicators. This requires time and investment.

    Ensure data compliance

    All data collected must meet necessary regulations. For consumer applications, GDPR requires user permissions and anonymisation. The EU’s Clinical Trials Regulation have restrictions around how organisations record, process, store and handle data, accurate reporting and verification, and confidentiality.

    Ensuring AI algorithms themselves are compliant is increasingly important. In April 2019, the FDA released a discussion paper on an AI framework for approving medical devices that use AI. It proposes best practices and expects manufacturers to provide transparency and real-world monitoring of AI devices, as well as updates as algorithms evolve. Expect this to become a rapidly evolving area of regulation which requires close attention.


    Use the right tools for the job

    Armed with the right data, organisations need to secure the right tools. These could be deep learning or machine learning algorithms, or statistical models. There is no single answer; an experienced data scientist will have seen enough to know what options are best for the job, whether it is adapting previous algorithms or creating new ones. They’ll also need suitable platforms to deal with lots of data. These should be decided based on needs, not on which tech provider has the best marketing campaign.

    Establish causal connections

    When building AI algorithms for the healthcare sector, correlations between data and outcomes will never be enough. There needs to be a causal connection before we can trust an AI to make decisions. This can be done. You could compare data to control groups in order to eliminate variables, or work closely with subject matter experts who can recognise driving factors through their understanding of the underlying science.

    Make it explainable

    Where necessary, AI should be designed to explain how the model reached its decision. If the user doesn’t understand what’s driving the AI decisions, they’ll struggle to trust the results. Explainable AI (XAI) involves designing systems with provision for interpretation and understanding of decisions. This is an actively developing research area.

    Validate your models

    Trusting that the AI works is essential, which means careful validation of algorithms before they’re used in the real world. This can involve various processes such as running the model on well understood new data to check it reaches the right conclusion, independently checking the code, or having medical experts assess the results and confirm they make sense. This is even more important where AIs are too complicated for full ‘explainability’ to be built in.

    Operationalise from the start

    Data scientists are getting pretty good at developing sophisticated AI models in the lab, but these often fail when deployed in a real world operational or consumer environment. AI is not just about the model. It must integrate with other IT or operational systems. It must be designed with its eventual deployment in mind, considering cloud architecture, devops, user interfaces, and more.

    Consult with both users and IT teams early on to understand these challenges.


    Focus on user experience

    Ensure your AI is designed to be easy to use, or people won’t use it.

    Healthcare product manufacturers are used to selling to doctors and expect them to just get on with it. With a growing expectation of easy-to-use technology this is no longer enough. We can’t expect people to engage with a second-rate user experience, and this is doubly true where the end user is a consumer.

    The ultimate aim of AI is for it to act as a collaborative partner to people, to inform decision making and generate trust in its data-driven conclusions.

    Deploy gradually and continually assess

    Some AI deployments will need to be done specifically to test them in real world healthcare environments in order to gather sufficient data to train and improve their accuracy.

    For example, Deepmind’s Streams, a tool to diagnose kidney disease, is being trialled by the UK’s National Health Service (NHS). This approach allows the AI to evolve in the real world, where doctors can flag when they disagree with a decision, which in turn helps the AI learn.

    It’ll take time for people to trust a machine with healthcare decisions

    It’ll take time for people – including professionals – to trust a machine with healthcare decisions. Don’t expect immediate results. AI systems need to be rolled out gradually with checks until they’re proven. For example, a doctor may start by making her own diagnosis, then run a diagnostics AI to validate it, or consider other options. Over time, she may trust the AI on its own and make diagnoses in parallel. Eventually, the AI may be the first port of call, with the doctor only brought in for serious cases.

    Expectations need to be managed as AI use grows. AI has huge potential, but it needs time to learn in real world environments to win over the sceptics. Overpromises early on can lead to long term suspicion, as IBM Watson found.