As a company, we have and continue to be committed to improving healthcare access, outcome, improvement and flow for patients through the use of data and advanced solutions. Over the last year, I admit that we have been preoccupied with the development of solutions to address the challenges in several key areas critical to progressing the state of Artificial Intelligence (AI) use within healthcare, one of the most prominent being clinical variation and how it affects patient outcomes, as well as their journey through the healthcare system. With the release of NHSX’s new report on the future of digital technology in supporting and improving an all-encompassing breadth of determinants to the smooth running of our health care systems and to patient safety, tackling this unwarranted variation has been highlighted as a powerful component of the NHS Long-Term Plan.
It signifies an important move in the right direction for the NHS and predictive analysis, building upon the NHS’s aim to be a leader of AI technology within 5 years. It is clear to see from the opportunities presented in the report that AI and its wider role in healthcare is now being considered much more seriously, particularly in terms of image recognition and algorithms to automatically process scans. However, the possible applications of Machine Learning (ML), I believe, exceed even this, and remain for us at Draper and Dash a somewhat untapped resource. The apparent push reported by NHSX to identify those truly utilising AI to the benefit of improving our health service vs. those who may be over-hyping the potential of their solutions is especially encouraging, and will be a good foundation upon which to allow real understanding of how far AI is spread across the healthcare sphere. Employing governance and a code of ethics for algorithm creation is a good start to this, ensuring the correct structures are in place to maintain a more sensible and regulated approach to embedding AI within the NHS.
Algorithmic explainability, or principle 7, proves a very important feature of this report, emphasising the need for ‘black box’ methods to be fully able to be backed with skilled data scientists who understand the mathematics and intuition behind the workings of algorithms used. It is necessary to consider, however, that certain more advanced methods described in the report, including convolutional neural networks to process diagnostic scans, would require simplification in order to aid understanding of the algorithm across the board. I am confident that such a process will lead to algorithmic explainability, organisation transparency, and shared understanding of the role of AI and how it can augment decision making. I use the term augment, of course, as no algorithm can replace the judgement and expertise of an experienced radiographer or clinician.
Another defining aspect of the report is tackling what ‘good’ innovation looks like and how AI, predominantly ML, can be used to tackle four key domains: precision medicine, genomics, image recognition, and operational efficiency. Deep learning models for detection of genomic sequences and a focus around how free text information captured in systems can be further processed by more advanced approaches through methods, such as Natural Language Processing, hold particular interest. For me, however, what needs to be considered alongside the rise of AI in healthcare is also the technology necessary to support these advances, in turn meaning more funding is required to better shape and prevent limitations on the future of AI.
To Draper and Dash, the workforce of the future section of the report truly resonates, directly aligning with our aim of helping healthcare providers plan better with the support of the right AI arsenal to do just that. It is clear and encouraging to me after reading NHSX’s report that there is a shared agenda between our work and that of others in the digital healthcare sector to adopt the use of AI within the NHS. With over 10 years of experience delivering award winning solutions, and having amassed a number of specialised ways that supervised ML can be adapted to predict key activity measures – such as readmissions, length of stay, diagnostic turnaround times, operation cancellations, and stranded/super stranded patients – there are many opportunities in which we as a company can help to shape the way machine learning is implemented into an NHS setting. There is further scope to extend our wider AI solutions into other areas, such as deep learning, natural language processing, image recognition, and help to build cases for new and innovative solutions, alongside a cultural paradigm shift towards our aim of augmented intelligence with AI/ML.
And so, the question at hand is this: How do we best deliver this clear need for utilising AI to its full capacity in order to support clinicians? Already, Draper and Dash functions in a space where we are utilising machine learning – deep, supervised, and unsupervised – to enhance many of our business intelligence solutions. With the release of this report comes a further opportunity to extend our product range by working with customers who might wish for something wider than our traditional machine learning methods. As such, we are constantly researching our potential to use the wider functionality of AI outside of ML, such as in image recognition, voice recognition, pattern detection, anomaly detection, and reinforcement learning among many others.
Since the outset of our work, we have strived to apply our specialist AI and ML platforms to driving the enhancement of existing healthcare systems. Recently, we have furthered our cause through the acquisition of Civil Eyes Research, a long-standing leader in providing benchmarking and healthcare insights for quality and productivity metrics to aid the understanding of health service functionality. As we combine our forces, we also combine collective years of expertise and crucial analytical capabilities in our commitment to making these important resources available and easily applicable to the healthcare team. As I mentioned at the start of this piece, much of my focus over the course of the past year has been upon tackling clinical variation. Kim Sutherland and Jean-Frederic Levesque discussed in their paper “Unwarranted clinical variation in health care: Definitions and proposal of an analytic framework” published just this year the existence of a rise in global interest in unwarranted clinical variation, but without the necessary methodology in place to tackle it – a strong reinforcement of our goal to address this gap in healthcare analytics and provide solutions to an issue that leaves few to no facets of healthcare untouched.
Confronting the subject of clinical variation is a vital step in optimising several areas, such as medicines and their impact on patient flow and cost. Medication is a ubiquitous component throughout almost every form of health care, and NHS RightCare has previously made clear the importance of reducing unwarranted variation and increasing value through optimisation of medicines. The issue arises when these medicines are used ineffectively, resulting in a damaging impact upon economy, society, healthcare systems, and most importantly, patients. While medicines comprised 47.6% of hospital spending in 2018, this level of fund allocation demonstrably does not translate to provision of the highest standards of patient care and outcomes. Through medicine optimisation, patients rightly become the focus of their medicine-based treatment plans to ensure they are fully supported in reaching the optimum outcome, with the added benefit to hospitals financially of reducing wasted medication spending. This is a responsibility that rests extensively upon the shoulders of physicians, but with the proper application of AI, those that hold this duty to prescribe in a way that best supports the patient’s health will gain better understanding of their efficacy, adherence, and effect on the patient. As Eric Topol examines in his recent book Deep Medicine, the current applications of AI in drug discovery – an invaluable tool to pharmaceutical companies – are becoming more and more widespread, but perhaps more applicably to the healthcare setting is its use in optimal drug dosage prediction. Due to the multifactorial nature of ascertaining the correct dosage, for example age, gender and genetics, the application of deep learning algorithms to more accurately determine just the right, personalised dose for each patient is a possibility that would once again greatly benefit both the patient in receiving their ideal care, as well as the clinician to better visualise and understand the patient’s needs with greater efficacy and efficiency.
Discussing the issue of applying AI to the clinical routine, Joerg Aumueller of Siemens Healthineers highlighted the possibilities in clinical pathway optimisation and refinement, “intelligently integrating longitudinal patient data and co-relat[ing] insights from imaging, pathology, lab and genetics” in order to reduce clinical variations and reduce diagnostic inaccuracies. Indeed, a pilot study undertaken at Flagler Hospital (Florida, US) earlier this year trialled the use of AI in reducing clinical variation in a way that “improves patient outcomes and is key to lowering health costs and succeeding with financial risk”. With clinical variation presenting as a costly and widespread issue comprising huge proportions of unnecessary healthcare, Chief Medical Information Officer of the hospital Michael Sanders reported the successful application of AI to “improve and standardise care paths for pneumonia, sepsis, COPD, heart failure and is on track to apply the same technology to 18 more conditions.” From a financial and patient flow aspect, this use of AI held the added benefit of saving $1,350 per patient, reducing length of stay by an average of 2 days, and reducing readmission rates from 2.9% to 0.4%, an encouraging example of how AI technology is already delivering on improving patient care, patient flow, and hospital costs in the real world.
In-line with the prevalence of artificial intelligence use in medical imaging described by the NHSX report, we have committed huge amounts of time and resources to support organisations through technology and transformation with clinical variation in areas such as radiology and the ordering of CT scans. Clinicians have been getting really creative in addressing the admissions avoidance challenges by getting more and more tests done at the front door, as the key concern here seems to be prolongation of a patient’s stay upon admission if they are then forced to wait for tests while in hospital. Even as the healthcare team work to improve the rates of patient flow, the issue of human error, especially under situations with tight time constraints, still remains. A number of commercialised medical algorithms – with little explanation of their workings – are already in use, with the additional checkpoint of a radiologist. However, as Eric Topol states in his book, “what if the radiologist is rushed, distracted, or complacent and skips that oversight, and an adverse patient outcome result?”. Hosny et al. further pinpointed radiology as a target for AI as a result of the need for improved efficacy and efficiency in their 2018 paper. Their work points out the continuous surge of imaging data that greatly outpaces the training of new readers, resulting in growing workloads and a particularly disturbing statistic: “an average radiologist must interpret one image every 3-4 seconds in an 8-hour workday.” As we can all imagine, these are not the ideal conditions for optimal analysis and patient diagnosis. In response, opportunities for the application of machine learning, neural networks, and natural language processing are being devised.
For those of you who read what we have been doing in predicting stranded patients, this is both exciting and value adding for the organisations who are looking to get to an optimal flow of patients.
This work in fact provides a useful case study for current AI usage. Our stranded supervised ML module is a predictive solution that looks at patients early on in their inpatient journey and uses various ensemble supervised machine learning methods to train a model looking at patient demographics, treatment specific issues, and comorbidities, amongst other features. This model is then tested, validated and pushed into a production environment to make an estimate of whether a patient will become stranded/super stranded. By estimating whether a patient will be stranded very early on in their journey, our algorithm allows the service to plan for a longer length of stay and to gear up the inpatient bed for this stay. What the algorithm also does is take into account which of the predictive dominant factors most effects that specific patient, allowing for a multi-disciplinary intervention to take place, if necessary, such as engaging social care. This predictive algorithm is then embedded into our products to allow the healthcare service to look at designing the system better for these types of patients. The core output and deliverable to be taken from this is to allow the services to respond to specific patient cohorts more quickly and more effectively. And we are still doing so much more with ML and early stage AI in healthcare.
With the many applications of AI in optimising medicine prescribing, patient care, flow and outcomes, as well as functionality of the NHS itself, there is one aspect that may be overlooked but should ultimately be at the heart of any methodology relating to the healthcare system: the doctor-patient relationship. Eric Topol’s book Deep Medicine makes the important point as to the ultimate focus of the health service, and what we should hope to achieve through our combined efforts in establishing AI methodology to lessen the burden of data collection, analysis and interpretation on clinicians. With this, the author hopes to reverse the current state of patient care in which doctors lack the time to be able to fully connect with their patients, freeing them to “restore the care in healthcare.”