Why are the expectations for AI solutions in healthcare sky high, but only few of them received authorisation and even fewer are used in the clinical setting? The implementation of AI in clinical practice is extremely challenging. In this fifth blog in a series of AI we discuss the 8 biggest challenges and recommendations in data organisation and integration of solutions into an existing workflow.
AI Challenge: Data
The most crucial and challenging step for developing and training AI solutions is having large amounts of compatible high-quality data that guarantees privacy and security. For healthcare specifically this comes with several challenges. Currently, there are still countless clinical processes, patient data and patient reports that have not been digitalised. Those that are digitialised, often are unintegrated. We have defined the 8 biggest challenges in implementing AI in clinical setting:
The most crucial and challenging step for developing and training AI solutions is having large amounts of compatible high-quality data that guarantees privacy and security.
1. Quality of the data
One of the risks of applying AI to clinical issues specifically is discrimination and introduction of bias. The quality of algorithms highly depends on the quality of the population it is trained and tested on, which needs to be representative. Issues will rise to any sub-population under-represented in the training data, as the AI model performs better on the group used to train it.
2. Ambiguity and lack of structure and interoperability
Another issue is that medical information is presented in unstructured files with a lot of ambiguity. Finding the right resources for widespread data collection and merging data sources remains a big AI challenge. We need to find methods to minimise missing data, to sample multiple providers, facilities, and patients, and to include evaluation of potential confounding variables to minimise bias and allow generalisation to broad populations.
3. No loop of real-time data
To truly use intelligence of machines, a connection must be made between real-time data gathered from patients and a central location where other datasets, for example of specific cardiac conditions. We need mobile devices transmitting patient metrics to deliver clinically accurate, medically verifiable data upon which to base AI’s algorithms. Otherwise, the ensuing analytics results are worthless.
4. Privacy and data security
With the new GDPR, handling data anonymously has become key. In many instances, patients and employees need to give permission for their data to be used. There are also security issues.
Implementing AI into the medical workflow: our recommendations
Another big AI challenge is implementing AI solutions into existing medical workflows. We propose a framework (fig. 1) that will guide stakeholders with the implementation of AI solutions in healthcare.
5. Creating support & algorithmic explainability
To create support, the interests, concerns and consequences of a certain AI application must be thoroughly mapped for each of the stakeholders. All stakeholders must be educated on the potential beneﬁts and risks of using medical data in AI techniques. This ‘algorithmic explainability’ is key because the ‘black box’ nature of digital healthcare is a thread for its trust and acceptance
6. Finding the right AI partner and techniques
Choosing the right pioneering partner for your AI solution is key in setting up the entire architecture from data flow to application management, thereby offering long-term commitment whilst at the same time having expertise in the field.
7. Plan, test and implement
After the right tool and AI partner have been chosen, it is key to have a decent action plan for its implementation. It is key to allocate a project manager, have a multi-disciplinary team of employees involved and ask users and patients for their feedback.
8. Continuously optimising use
The process doesn’t end when an AI tool is developed and implemented. Any AI tool should fit perfectly into the current workflow. An experienced and communicative data scientist – allocated by the AI partner or internally – is needed to make sure AI tools are streamlined and optimised.
Fig. 1. Framework for implementing AI
Read our whitepaper
In our whitepaper we discuss the basic workflow of AI tools, we provide an overview of the opportunities and AI challenges of implementing AI in healthcare and give actionable insights to help you overcome barriers and develop smarter strategies that accelerate the introduction of AI innovations.
Download whitepaper AI
In this whitepaper ttopstart will give you insights on accelerating Artificial Intelligence for healthcare and life sciences. This paper will provide an overview of the opportunities and challenges of implementing AI in healthcare and give actionable insights to help the different stakeholders to overcome barriers and develop smarter strategies that accelerate the introduction of such AI innovations.