The (not-so-distant) future of AI in healthcare

There have been many claims about technologies making healthcare better or more affordable. Artificial intelligence (AI) is one of the few innovations that can deliver on both these promises.

In this series of articles, I look at how AI can improve healthcare systems, without compromising on quality and costs. First, I review the impact of currently available AI applications. Then, I outline the most promising use cases for the near future. The series ends with ways to build a business case for adopting AI.

This second post is my vision of the near future of medical imaging AI. It includes insights from conversations I’ve been having with two experts in the field.

The “health bay” versus current cancer care

Set in 2154, the sci-fi film Elysium envisioned a luxurious health facility in which a device, called the ‘health bay’, could detect and heal all medical problems within seconds: instant diagnosis, treatment recommendation, and treatment, in one location. We are nowhere near any of these capabilities, but it is a vision to which we aspire.

Our society is witnessing a growing and ageing population, increasing the pressure on healthcare systems at a time when budgets are limited, and the workforce is insufficient. The care pathway for cancer, the second leading cause of global deaths, sees additional challenges. Oncology uses a multitude of diagnosis and treatment modalities, involving different medical centres and following increasingly complex guidelines. Advances in diagnostics and treatments are welcome, yet they also result in more clinical parameters to be analysed for each patient. 

Collecting and analysing patient data stored in different systems, e.g. medical scans, tissue samples, or lab results, in time for a multidisciplinary meeting (or MDT, a discussion involving multiple physicians, part of the cancer care standard), is a manual, time-consuming task. Moreover, humans are not well-equipped to analyse large volumes of complex (medical) data and will, at times, make mistakes.

AI today is already making an impact in the detection stage of the oncology pathway, by quantifying findings on medical images. Consequently, AI tools are already collecting data that can improve the next stages of the pathway. However, the current systems are still ‘narrow’. The AI analysing a medical scan knows nothing about the patient history. Its analysis is restricted to abnormalities, and the system cannot detect multiple diseases accurately.

Dr Rizwan Malik, Consultant Radiologist at Royal Bolton Hospital, Director of South Manchester Radiology, notices the limited focus of current AI solutions:

“Far too much attention is on the narrow ‘detect abnormality on images’ part of the pathway to the detriment of all the other stages with so much ‘low hanging fruit’. Why? Because it hasn’t seemed ‘sexy’ enough. As I often say: boring AI is good AI.“

Artificial intelligence has the potential to support or realise significant improvements in cancer care. Seeing the immense gap between where we are today and where we wish to go to reduce cancer mortality, there is an urgency to make that happen.

Integrated diagnostics with AI

Solving the challenges above in the not-so-distant future takes two switches: aggregating more data and building and integrating multi-modality AI systems. In Dr Malik’s words:

“We chase and crave digitisation to make health data more accessible and then lock it up into disparate digital silos. (…) Being able to access, aggregate, analyse and then implement all these pieces of data should be leveraged throughout the oncology pathway.”

To take the next steps, we must first define and create an AI-powered platform to automatically collect and integrate relevant diagnostic data. The data would come from different modalities that are part of the oncology pathway. For example, it would consist of medical images, tissue pathology, molecular biomarkers, genetics, and patient medical records. Important to emphasise here is that there is no need for any patient-identifiable information as part of this data.

The platform would function as a database on which AI systems can further be trained to provide more input for diagnostic decision-making – a multi-modality AI, as displayed in the image below. One or several AI algorithms developed would analyse the data at different stages of the oncology pathway: screening, detection, staging, treatment recommendation and response assessment.

Bringing all this information together and analysing it with AI creates a more precise overview of a person’s disease, useful as a dashboard to support decision-making during MDTs. In these meetings, clinicians match the results with novel treatments or trials to provide the best follow-up.

The benefits of AI-enabled integrated diagnostics would be considerable. We’d see more efficient reporting, faster and more precise decision-making, and increased accuracy in assessing treatment response. In essence, we could have more tailored treatments and deliver better patient outcomes.

The lung cancer example

There are several clinical needs in the lung cancer pathway. For each, I propose ways in which data collection and AI solutions may help.

1. Early detection

As Aidence, we have been working on improving early lung cancer detection and reporting with Veye Lung Nodules and Veye Reporting. Our AI solutions analyse thousands of chest CTs each week to help radiologists manage lung nodules, in both routine practice and lung cancer screening. This stage is where we are now, focusing on driving further adoption.

2. Diagnosis 

Determining the best diagnosis and follow-up for a lung cancer patient is a difficult decision, requiring a lot of data. Based on clinical needs and AI capabilities, there are several use cases to investigate.

A first assumption to explore is developing AI to predict malignancy. In a 2017 Kaggle challenge, we built an algorithm that was successful in predicting lung cancer on a single scan. AI could also play a role in classifying the extent of the spread of cancer by advising the TNM classification of malignant tumours on PET scans.

Secondly, there is research showing that AI can predict recurrence based on tissue samples or CT image analysis. We hypothesise that you can bring both together and get a better predictive model for recurrence risk prediction. We are currently exploring this hypothesis with Algnostics. 

Thirdly, there is a need for a data-driven recommendation of a patient’s need for biopsy, surgery, or a specific type of therapy, which an AI-enabled analysis can provide. For example, there are cases where experts disagree on the need or results of a biopsy, an invasive and costly intervention. Reducing the number of unnecessary biopsies would benefit patients and healthcare systems.

3. Treatment response assessment 

To assess treatment response, radiologists are currently looking at the evolution of lesions in follow-up CT scans. AI could automatically measure nodules, lymph nodes and masses, and compare them with multiple prior scans to determine changes. Physicians would then have more time to interpret the information.

AI can take the first step in making tumour assessments more accurate. A current set of guidelines on tumour measurements, RECIST, are based on diameters: a certain percentage of diameter change indicates if the cancer is growing. AI systems can already provide volume measurements. Volume doubling time, for instance, is an internationally recognised parameter for the growth or decline of a lesion. Using AI presents the opportunity to make it a part of the guidelines.

4. Data aggregation

Data is being generated across all stages of the lung cancer pathway, supporting patient management and reporting. We could make a better, more intense use of this data.

For instance, our interactive application Veye Reporting produces standardised reports following the protocols of the Targeted Lung Health Checks in the UK. As the data produced is standardised and delivered according to specific protocols, it can be used to create new insights for quality assurance or improvement of the lung cancer screening programme. Case in point, if we analysed reports collected across different regions, we could define a benchmark for recall rates and highlight areas where this rate is too high.

5. Intelligent automation of operational processes

One of the least ‘sexy’ applications for AI is the automation of administrative and operational processes. Think about the interpretation of medical notes, radiology reports, and patient and population management.

For example, effective lung cancer screening requires GPs to identify at-risk patients in their populations. This is a manual search on specific characteristics in the patient records. AI could make this process faster and more precise through automated processing and by flagging patients for yearly follow-up. Similarly, based on clinical guidelines, AI can analyse if the patients’ diagnostic data is complete for discussion in MDTs. The benefit is preventing unnecessary scheduling of patients.

One step further: prognostics and new standards of care

Looking ahead to the next three-five years of developments in healthcare AI, changes remain in line with the current clinical protocols. These standards of care rely on human knowledge, which is considerable,  yet far from complete on all diseases. Thus, there is room to make them better.

Farther down the line, AI findings can contribute to new standards of care, tailored to individuals rather than to groups. AI systems can compare the data of a patient with that of thousands or hundreds of thousands of other patients with a similar illness. Then, it can predict the chance of a patient developing a condition, or responding to a specific treatment.

Physicians would still check and confirm any new clinical protocols. AI is far from an autonomous system, and we still question whether that is desirable in healthcare.

An industry shift

Building an AI-integrated diagnostics solution for the entire oncology pathway is a daunting project. It requires years of development and collaboration. It is unlikely that only one company will develop all the AI algorithms. An AI-enabled pathway requires a partnership between academics, focused AI start-ups, pathology and imaging experts, providers, and so on.

Kicky van Leeuwen, AI in Healthcare Researcher at Radboud UMC in the Netherlands, on an industry shift:

“When AI will become more prevalent in the clinic, it will be more difficult for departments to maintain the different contracts and infrastructures. Therefore, we see the trend of intermediate platforms and marketplaces evolving, both from new parties as from the ones we know.” 

We have seen initiatives fail due to complexity. So, I expect a few specialised companies that will integrate their solutions into an infrastructure that covers multiple aspects of a pathway. Eventually, mergers and acquisitions may create a full pathway offering.

Toward excellent care

There are good reasons to explore ways to improve healthcare. AI certainly holds great promise for future use cases across the oncology pathway. Dr Malik summarises the potential of AI to make high standards of care accessible to all patients:

“The aim for any AI application within oncology should be to assist in arriving at an outcome more quickly, (…) while at least maintaining (or indeed improving) the quality of diagnosis and service. This harmonisation of excellent care should be decoupled from where it is delivered. (…) Patients should be able to expect the same high standards irrespective of where they are seen.”

For all these efforts to be successful, the AI medical device industry has an important responsibility: upholding high-security standards for data collection and processing. As Aidence, we have set a solid foundation on which to expand our suite of clinical applications for better cancer care.