A Different Ball Game - AI in Medical Diagnostics
- Dr Rob Porter
- Jul 11
- 6 min read
There have been a lot of articles in the news recently about AI, where ‘prospective’ Wimbledon line judges were protesting about Hawk-Eye technology making their role redundant. Whatever your opinion about replacing people with machines, no one will suffer a serious injury for an incorrect line-call decision. This was proved when someone accidentally turned Hawk-Eye off mid-game – a player got a bit annoyed, a commentator was confused, no one died. AI in medical devices is a different ball game.

Driverless cars on the other hand, carry a lot more risk and some ethical conundrums. Another article posed the question that if a car’s AI calculates that, irrespective of the available actions in a certain situation, one of several people might die, how should the AI decide who lives and who dies? Driverless cars scare me, partly because I’m more comfortable being in control and partly because I’m deeply uncomfortable with AI making life-altering ethical decisions.
These news items prompted me to reflect on what I’ve learned so far regarding the various aspects of AI in medical diagnostics. I am certain that AI can provide a great service and should be employed in more areas of diagnostics. However, the examples above are comparable to the potential issues that AI faces in medical devices.
I know that there is work being done to improve software and firmware to provide better reading, analysis and prediction through algorithms and learning. We have to be acutely aware of the risks associated with technology failing and this resulting in patient harm. AI in medical devices can be very risky and contrary to some software programmers and APP developers’ beliefs, far from simple to implement. Some are unaware of the risk factors and others have never heard of ISO62304 – which is the standard required for CE, UKCA and MDSAP regulatory approval. FDA have their own standards, but a lack of knowledge of ISO62304 is a pretty solid indicator of FDA standard ignorance too. A developer can offer outstanding AI programming and design skills, but without a deep understanding of the factors at play in the diagnostics industry, their output will not reach the market.
ISO62304 is the standard for ‘Software Lifecycle’ and is mandatory for any software under the harmonised standards of ISO13485 framework. Whilst this standard provides a great framework for software and AI development, its implementation is very challenging.
Using Lateral flow devices (LFDs) as an example, as they have been a cornerstone of rapid, point-of-care diagnostics for many decades (as have glucose sensors and other electroanalytical/chemical, optical or magnetic mediated devices). This platform has been implemented firstly in home pregnancy tests (Unipath’s ClearBlue device) to infectious disease screening (the home COVID19 test). These simple, portable tools have made health monitoring more accessible than ever before. Yet, as technology surges forward, a new frontier is emerging where AI converges with lateral flow technology, unlocking new levels of sensitivity, precision, and connectivity. So how can the role of AI be pivotal in transforming LFDs, exploring the science, applications, challenges, and future prospects.
AI, particularly in the form of machine learning and computer vision, is emerging as a powerful ally to LFDs. AI’s capability to process, interpret, and learn from complex data patterns, therefore making it potentially ideally suited to address the shortcomings of traditional lateral flow testing, where the user’s ability to correctly identify the true result is being compromised by a preconception of a desired result.
Key benefits of integrating AI with LFDs include:
· Objective Interpretation: AI algorithms analyse test strip images and determine results with greater consistency and remove user bias.
· Enhanced Sensitivity and Specificity: Advanced image analysis can detect faint lines or subtle changes, improving accuracy in borderline cases.
· Quantitative Analysis: AI can estimate analyte concentrations, turning qualitative LFDs into quantitative diagnostic tools. Something I did with the MyLotus system.
· Data Connectivity: Digital results can be instantly shared with healthcare providers, facilitating telemedicine and epidemiological monitoring. This though has its complexities.
· User Guidance: AI-driven apps can instruct users in real time, reducing errors in sample application and interpretation.
Now before we go any further, I want to talk about Firmware. You have Software and Firmware, the Firmware is software code, which controls the Hardware. Then on top of that you have the operating application software and then the GUI (Graphical User Interface). So why talk about Firmware, the application software does all the algorithms calculations and work surely? Well yes, but if the Firmware changes it can impact the application software. This is ok for dedicated readers where the Firmware and Hardware can be locked. Mobile devices on the other hand, are different, APPs rely on the operating system (the Firmware of the device) and the Hardware (the camera quality and lighting).
APPs are therefore a complex issue. Firstly, you have different operating systems (most popularly, Apple IOS and Android) and within these platforms you have different versions of each operating system circulating in the market at any one time. Add to that, different hardware, as cameras and lighting differ in each model. Then horror of horrors, we have the phones AI kicking in trying to improve the aesthetics of the picture, changing colour balance and contrast to make it look appealing and not allowing us to get to the raw data.
I’m not anti-APPs and AI. These issues can be countered, but it’s not as easy as writing the software and off you go. There are many variables to consider. This maybe why Medical APPs have really not taken off in the same way that wellness indicators such as heart rate and stress apps have on phones and smart watches. Many wearables aren’t medical devices because they are not used for proper medical analysis (not my opinion). Whilst these APPs sit in a grey area, they create the perception of being easy to transfer to medical diagnostics, as (to all intents and purposes) they look medical.
So, what things can be done to make things easier (this is no means exhaustive, but a snapshot of some mitigations)?
· Firmware changes: Perform all measurements and calculations in the cloud, where firmware can be managed centrally. The app then simply displays the results, making it easier to control different system and versions across mobile devices.
· Hardware changes and mobile device AI photo enhancers: Here the manufacturers need to be helpful (easier said than done). When photography is a major selling point, manufacturers are understandably very protective of allowing people control and access to their AI IP. Manufacturers could however, enable developers to disable AI features and control camera pixel ratings. This would simplify analysis and improve information prediction by enabling learning from historical data.
· Lighting: Light sources vary in colour and intensity. To address this in quantitative measurements, a reference mark with known optical density is used so software can adjust and correct for lighting differences.
· Change controls: Tools such as GIT programs are available to monitor and track software changes. For major modifications in code, it is recommended to provide clear explanations. Code should not be released without appropriate documentation of significant changes.
Finally, there is the issue of how medical diagnostics are used within a point-of-care system. Writing independent software and APPs for healthcare professionals may be designed to save time and improve convenience, but this isn’t necessarily the case. Hospitals need to comply with regulations like HIPAA, GDPR, and SOC 2 which add complexity to any software and data integration. Consequently, many hospitals still rely on tried-and-tested (outdated) systems that are not designed for modern data integration.
Healthcare data is highly sensitive and any breach can have serious consequences for patients and the organisation. Any changes to the operating system would require risk assessments, mitigation and testing plans to ensure all the software in the hospital is still function as it should.
Introducing new systems that access sensitive patient data during integration requires robust encryption, access controls, and user management across integrated systems. Data integration solutions must be designed with the needs of end-users in mind to ensure they are used effectively. Integrating data in real-time, can be complex and resource intensive. Healthcare organisations need skilled personnel with expertise in data integration and ongoing software maintenance and support of integrated systems. It’s not a commitment to be taken lightly – and can be a drain on limited resources.
These are a few of the considerations that make a medical device APP uniquely complex, expensive and time consuming. They are not the sort of technology that can be delivered by someone with ‘some coding experience’.
The marriage of AI and lateral flow technology is still in its early days, but the trajectory is clear: faster, smarter, and more connected diagnostics are needed and on the horizon.
Key trends shaping the future include:
· Personalised Medicine: AI-enabled LFDs could tailor diagnostics and treatment recommendations to individual users based on longitudinal health data.
· Multiplex Testing: Devices that simultaneously detect multiple analytes and interpret complex patterns using AI will provide comprehensive health snapshots from a single sample.
· Global Disease Surveillance: Real-time aggregation of anonymised test data empowers public health agencies to anticipate and respond to outbreaks before they escalate by predicting disease spread.
The pace of change in the world of AI is so fast, that any prediction of its future impact on IVD technology could be obsolete in a matter of weeks. So, in the tradition of the movies based on the perils of super intelligent computing – I’ll keep some opinions and thoughts in reserve – for a sequel.
You lucky things.
Rob
Comments