Artificial Intelligence For Healthcare: Potential and Pitfalls

Artificial intelligence (AI) is often perceived in media as evil robots trying to take over the world, but what If I told you AI could save lives, and improve healthcare for millions of people around the world? With organisations being data-driven and dependent on modern technologies to get the most out of their assets, it’s no surprise that the healthcare industry is keen to follow suit.

I was recently fortunate enough to represent Capgemini at the 5th annual ‘AI for Health Summit’ in Paris. Here, I had the opportunity to engage with experts across the medical, pharmaceutical and technological industries. In this article, I will look at what artificial intelligence is, how AI and data are changing the scope of the healthcare industry; identify the current roadblocks and discuss how leaders in the 'Med Tech' industry are using innovation to overcome them.

What is AI?

Artificial intelligence is the development of software and computational systems to mimic human intelligence and perform tasks. Machine Learning (ML) is the subset of AI that has the ability to learn automatically from a dataset, without the need to be explicitly programmed. It builds upon its ability to understand and recognise patterns, improving its decision making and the formulation of accurate predictions.

Potential for AI in Healthcare

How does all of this link to the Healthcare industry? Well in simple terms, faster medical treatment saves lives. In many pressing medical problems, the answers to knowing who to treat, when to treat and how to treat may already be in your data. AI is being used throughout the patient journey, from providing accurate diagnosis, to personalising treatments based on your genetic make-up. No two people in the world are alike but AI models are helping doctors learn from patients with similar conditions, or even genome. Healthcare professionals can then make highly informed decisions around diagnosis and treatment options.

As an example, let’s look at diagnosing skin cancer. This can be extremely complicated for the doctors in making decisions and diagnosis, for the patients in the understanding of the risks and the success rates of available treatment options. Researchers at Max Kelsen, an Australian based software engineering company, are developing an AI model to streamline the process of cancer diagnosis[1]. The AI model is 'trained' from a range of sources, including blood test results, X-ray imaging, clinical assessments and genetic information. It can rapidly consolidate this information and provide highly accurate predictions of the patient’s diagnosis, as well as treatment options with the highest rate of success.

Another example of AI potential is macular-degeneration and blindness. Studies show that nearly 10% of all clinical appointments in the NHS are for eye related diseases, roughly 10 million per year. With an ever-growing (and aging) population, there’s insufficient healthcare professionals available to properly deal. As a result, there will be a percentage of patients that will go blind/have their condition worsen due to a delay in response. In 2018, Moorfields and DeepMind published a paper describing an AI model that, given a retina scan, could make appropriate/correct(?) referral decisions 94% of the time, matching that of human experts[2]. The model can detect over 50 types of macular degeneration as well as any doctor and in a fraction of the time. It analyses retinal scans of the patient and within seconds, can delineate all of the features that correspond with a disease. For a human expert, this exercise could take hours.

In 2020, Deloitte published a report on ‘The Socio-Economic Impact of AI in Healthcare’. Their research showed that our current AI model can mature by collecting data from the population using methods such as wearable devices, imaging, monitoring and other real-world data. It has been calculated that, these AI-generated insights could save 400,000 lives, 200 billion euros and 1.8 billion working hours. All of this annually and just in the European Union alone.

Roadblocks

So, if AI can save time, money and most importantly, lives, then surely it needs to be implemented in every hospital available… Well, despite being a high priority for 75% of CEO’s and executives, only 20% appear mildly acculturated to data and AI[1] – so why is this? Although the potential is great, there are ongoing issues and pitfalls which we must acknowledge.

Inefficiency

One significant data challenge is the inefficiency in training AI models. In order to train a machine to identify and recognise patterns with a low error rate, tens-of-thousands, if not millions of data samples must be fed in. This means they’re inefficient in the learning process. For example, a human could look at the image of two woodpeckers below and clearly see that they’re different. Taking this a stage further, most people would be able to identify the differences in these two with a high certainty/low error rate. Now, to train a machine in order to achieve the same low error rate, a plethora of data must be available and inputted.

This inefficiency is a big challenge in healthcare, where we usually have very little data available.

Data Governance

Now of course we do have data as there are billions of records on patients and their medical history, from studies, registries and electronic health records – but we cannot just simply use all of it. This is due to data privacy laws, regulations and data governance. With more than 120 different countries already engaged in international privacy laws for data protection, access to this much needed information is becoming increasingly difficult. In recent years, DeepMind, Google’s AI arm faced legal action over NHS data use[1], having legally inappropriate access to over 1.6 million patients. The firm insisted that patients’ records were being used to create an application that could be used to alert, diagnose and detect patients who were at risk of developing acute kidney illness. However, in 2017 the Information Commission found that the hospital had not done enough to protect the privacy of patients.

Notwithstanding these issues, innovation from start-ups is looking to overcome the problem of data governance and protection with the use of Data Anonymisation. Upon my recent trip to the AI for Health Summit, I engaged in a demonstration held by Octopize MD about how they’re attempting to tackle this issue of data governance. Octopize MD is a French start-up that developed what’s known as the ‘Avatar technique’, a unique personal data anonymisation solution. It aims to solve the paradox between protecting patients’ personal data and the sharing of this data for its informative value. Since the establishment of the GDPR (General Data Protection Regulation), transferring personal data to other countries has been a blocking point for many projects. Thanks to its Avatar method, Octopize is able to create a synthetic dataset that protects the individuals at the origin and source, while keeping the statistical value, potential and original granularity. Their transformation of information is accompanied by an evaluation of synthetic data through unique metrics, developed to comply with the 3 criteria set out by the European Data Protection Board[2]; singling out, linkability and interference. From their transformation of data into Avatars, it is possible to accelerate and facilitate the transfer outside of the EU (theoretically can be applied to places like USA / Asia), while respecting the privacy of individuals and ensuring a strong preservation of the statistical qualities of the original information.

Data Bias

Providing you have the data available to train an AI model, then surely this should be okay? Well, not quite. Another issue many organisations face is inconsistency in their data, otherwise known as data bias. If we refer back to our skin cancer example from earlier, we must acknowledge the disproportionate representation of those from different ethnicities. Skin cancer affects white people at a higher rate, this is due to the lower quantity of melanin in the skin, which provides protection against UV light - the leading cause of skin cancer. Although it is not the only cause for skin cancer, by contrast, darker skin tones with more melanin filter at least twice as much UV light[3]. As a result, melanoma skin cancer is roughly 20 times more common in white people than black people[4] and on top of this survival rates are also shown to be affected based on your ethnicity. So how does this all fit in with AI? Well, when it comes to training an AI model to detect skin cancer, if the data we provide is comprised by a majority of Caucasian / lighter skin tones, how accurate will the AI model be in detecting skin cancer among those who are black / darker skin tones? When it comes to detection, cancer in lighter skin tones may be red, brown, or black. In darker skin tones, it may be similar to the surrounding skin, or darker. This makes it more difficult to detect and therefore the model has the potential for inaccuracies.

Speaking with Dr Yan Liu, Chief Medical Officer of Median Technologies, she explained that we need to account for variables in the data, not mitigate those that don’t match a particular category. Median Technologies is a Med Tech start-up looking to provide innovative imaging solutions and services to advance healthcare. Their aim is to increase the accuracy of diagnosis and treatment of many cancers. Dr Liu acknowledged data bias is a common roadblock, however explained how her team are overcoming sampling bias. It is key to ensure sufficient data is collected from a range of sources and locations. Utilising key data and insights from all across the world – not just a discrete sample such as the West, or East, but a combination. By doing so, you start to gain an equal representation of the human population, not a skewed dataset. One approach to solve this is extract, load and transform. Extract the data from different sources, load into a single repository, and then merge or transform the data. With regards to detecting cancer, you can use information from blood tests, in conjunction with imaging and consolidate this data into a single, unique dataset for analysis. This helps remove bias from variables such as ethnicity.

Data Quality

Data quality can be a challenge when it comes to an organisations decision-making ability. Inaccurate, inconsistent, missing and duplicate information poses a threat to cultivating trustworthy data sets; sets which we use to train our AI models. Gartner estimated that poor data quality directly costs the average organisation $12.9 million a year, with data professionals spending more and more time checking quality.

When it comes to data quality, there are issues that are distinctly technical or non-technical, and there are others which can be a result of both.

Duplicate data

A complication which can be a result of either technical or non-technical issues is duplicate data. This can occur for a number of reasons, especially in the healthcare industry. When working in a fast-paced environment, with hundreds, if not thousands of patients being treated each day, human error may occur when logging a patient’s information. Instances such as misspelling one’s name or details may result in multiple records for the same individual. This is an example of a non-technical issue with data quality. Technical problems such as pipeline errors can also result in duplicate data. If there’s a latency / network issue with the underlying infrastructure, health records may be read twice and both versions could be saved, resulting in multiple logs for the same patient.

Duplicate data not only results in inaccurate reporting and service problems but additional costs of storing duplicate records also needs to be considered. Fortunately, finding duplicate data is a relatively simple task to automate and resolve, than other data quality issues. If records are identical, or near enough, these can be flagged and reviewed quite quickly.

Incomplete data

Data completeness, whether it’s incomplete or obsolete information, is a huge problem when it comes to?healthcare. It’s difficult to be accurate in your analysis and be fully confident in the insights when information is missing.

When overcoming missing data, it is important to first identify whether or not the data is completely missing at random, or if there is a noticeable trend in the way it's missing. From here, there are multiple techniques to rectify the issue outside of contacting the patients individually to confirm their information. Research from Selerity, an organisation that offers statistical software services, highlights the importance of incomplete data and explains the two most common approaches to overcome it.

Deletion

These methods work best for datasets where a subset of values is missing (i.e., patient data has some information but not all). Using common methods involve deleting entries with missing values. This is advantageous when there is a large volume of data, as it will be less likely to distort the readings. Alternatively, pairwise deletion eliminates the need for certain variables that are missing, providing the variable is vital for testing.

Imputation

It is common for data scientists to use two data imputation techniques to handle missing information. 'Average Imputation' uses the average data from other entries to fill out missing values (important to note that this can artificially reduce the variability of the dataset). Alternatively, 'Common-point Imputation' is when you utilise the middle point (median) or most common value (mode) as a replacement for the missing data.

Missing data is sometimes an unavoidable situation and a frustrating part of the process when it comes to working with AI and building models. Using the above methods can help mitigate the effects of incomplete data, albeit not a permanent solution. The best means of handling situations such as this, is to ensure you have contingency plans in place to reduce damage.

Lack of Strong Governance and Quality Assurance

Finally, a lack of strong governance and quality assurance results in complications in an abundance of areas of a business, including the data and decision making that underpins them. A strong governance framework in any industry allows for a clear strategy to be developed and delivered. When working with clients, efforts must be made to ensure contractual agreements are met, with clear agreements on performance, design and other expectations. Meaning, if we have chief executives and senior stakeholders that aren’t implementing systematic practices, we will see an inconsistency from the top down. This inconsistency can be as subtle as a difference in naming conventions and definitions which lead to misinterpretations of data i.e., in different regions, different names may be given for the same object resulting in information being miscategorised. These nuances can result in huge discrepancies in data over time and ultimately affect any insights deducted.

Quality assurance is a crucial process and a cyclical method of assuring progress at the highest standards and quality. The process looks like this:

1.??????Identify issues with process

2.??????Generate corrective measures

3.??????Verify corrective actions

4.??????Implement corrective actions

5.??????Monitor and control

6.??????Repeat steps 1-5

This way of working allows us to ensure the AI models we train are as accurate as possible. By using this solution feedback loop, we can compare the AI model with results from real scenarios and analyse the two. Refining the model allows us to improve its accuracy in an iterative approach, where we can compare the two data sets and repeat the process as necessary.

Conclusion

The general consensus from experts across medical and technological professionals alike is vast excitement. AI is changing the way we work, and we will start to see a smarter, more agile and a safer healthcare system. It’s important to realise that although some jobs are disappearing to be replaced by smart machines, your doctor is not heading for an early retirement. AI will not take over the role of any medical professional, rather that AI will be used as a tool to improve the way they work. Albeit we’re still not ready for widespread use, the advancements in AI and technology are very promising. Establishing regulatory frameworks to ensure diverse and robust tools are developed, that are compliant and adaptive, and can serve the whole population equally. If we can get this right, we can transform the delivery of healthcare and help millions around the world.

[1] ‘Brisbane AI to help with cancer treatment in an Australian first’

[2] ‘Google DeepMind announces research project with NHS to detect eye diseases’

[3] ‘The AI For Health Book’

[4] ‘DeepMind faces legal action over NHS data use’

[5] ‘Article 29 Data Protection Working Party’

[6] ‘Skin cancer concerns In People of Colour: Risk Factors and Prevention’

[7] ‘Key Statistics for Melanoma Skin Cancer’