More than 80% of healthcare data is unstructured. Usually hastily typed by a busy clinician, it is full of professional jargon, shorthand, and frequent typos. Often, shorthand and jargon are specific to a department or hospital; sometimes the same phrase can have different meanings. For example, PT could be short for ‘patient’ in one orthopaedic department, and ‘physiotherapy’ in another, making it hard to understand by people and AI alike. To turn text into data, AI needs to overcome these challenges, and advances in AI increasingly help in this task. Recent advances in Large Language Model (LLM) technology, for example, have demonstrated that size can help overcome some of these challenges, but even they have an upper limit. As dozens of companies are eagerly exploring LLMs, Evidentli is already prepared for the next big evolution of artificial intelligence: Agile AI.
Why does AI need to be agile?
Agile Artificial Intelligence (AAI) stems from the observation that sustained and rapid advances are the only constants in the field of AI. Breakthrough technologies like Deep Learning and LLM provide giant leaps forward, which are normally followed by periods of incremental growth as the technology permeates across the industry and new applications for the technology are tested. Rather than integrating one technology at a time, AAI provides a dynamic ecosystem for hundreds of collaborating AIs.
Four main benefits of AAI are:
Underpinning AAI is the recognition that different kinds of data require different processing. In medicine, the same data can require different algorithms at different parts of the value chain and insights from the data can have a profound impact on how that data is used downstream. Analogous to how different specialists can look at the same patient’s health records and form different views of the patient, different kinds of AI can produce complementary insights from the same data. Agility means that AIs are created and trained based on their relative strength areas and the suitability of these strengths for each data segment. Different AI models can collaborate when they are controlled by an agile process. Collaboration can include co-learning - where one AI model creates and trains others forming a cascade of highly precise understanding of the data.
A key requirement for medicine is incredibly high precision, which remains a challenge for all AI. Performance that qualifies as “good enough” in other industries is usually unacceptable in medicine due to the effects of errors on human lives. Just like medical teams rely on specialists from different disciplines to provide coordinated care, AI needs to process data from different perspectives and determine the right actions that would lead to the optimal outcome. The cost of error is much higher in medicine than in other disciplines due to its direct and immediate impact on human lives.
3. Handling complexity
The data value chain in medicine provides a complex web with often competing goals and views. Interpretation of the data is multi-faceted depending on where in the value chain it takes place. Some insights can be produced by neural networks, while others require background knowledge outside the training data. One end of the value chain is focused on discovery and understanding the basic building blocks of data. At the other end, evidence is produced and so needs very different kinds of intelligence. Agile AI is required to be able to handle the entire value chain.
To be useful, evidence needs to be trusted. Trust comes from ethical, transparent and reproducible explanations of the data and how they were used to produce evidence. In complex medical value chains, this translates to a requirement that every AI along the way is explained. Such documentation is predicated on including the entire pipeline, which is not a suitable task for individual components in the pipeline. Agility means that as the processing pipeline changes in response to changes in data and requirements, its view of the entire process remains accurate.
What is Agile AI?
Agile Artificial Intelligence controls, monitors and verifies multiple other AIs along a value chain. It is a combination of a large number of AI models, usually hundreds or thousands, each created and used on the segment of the data it is best suited for. Importantly, these models don't all have to belong to the same AI family. One LLM can be used to understand discharge notes and another can understand progress reports. A third may use insights from both to create a narrative of the entire patient journey. A fourth can verify the chronology of the journey and a fifth can describe the entire ensemble. In real-world data, thousands of AI would participate in this collaboration. Segmenting the data and delegating it to different AI makes data processing economical, and more importantly, precise.
To make data segmentation and the construction of a large number of models feasible, their creation needs to be automated and controlled with agility - as the composition and nature of participating AI changes, the controller needs to adapt as well . Using AI to provide control adds scalability and optimisation to what are otherwise static pipelines that require a lot of manual effort to maintain.
Just like the human brain that uses lobes to perform specialised tasks, AAI dynamically segments input data, and delegates different segments to the most suitable subsystem. A subsystem can consist of one or more algorithms. The suitability of each subsystem to the task it is given is verified and adjusted as needed during a training phase. In this way, AAI autonomously adapts to the specific dataset it encounters. The result is that medical data is processed by AAI with precision up to 20 times better, not worse, than human domain experts.
Ideally, AAI uses “active” learners, meaning that they continue to learn incrementally through observation of other parts of the AAI and from human operators. Importantly, this usually includes the ability to ‘unlearn’ errors. AAI therefore needs to understand its constituent models and proactively manage their knowledge.
AAI is the key for healthcare
Medical data includes several challenges that are not currently addressed: a very low tolerance for inaccuracy, structural and semantic difference between siloed datasets, and sceptical adoption of new technologies. Traditionally, AI is not good at addressing these requirements because it requires datasets to be similar to each other, it is not transparent, and it has a relatively high error rate. AAI optimises the processing pipeline - as it sees more data, it examines different AIs combined in different architectures, dynamically adjusting itself to provide the best possible results for the data.
Evidentli: A pioneer of Agile AI
While more and more organisations are discovering and exploring AAI, Evidentli’s team with its decades of experience in all aspects of AI for healthcare, has been perfecting their Agile AI since 2010. Evidentli’s Agile AI consists of multiple generative AI, natural language processors, machine learning systems and more throughout the Piano platform. A typical ingestion workflow for real-world data (RWD) on Piano, sees hundreds of models being generated automatically, Agile AI is the key to Evidentli’s ability to handle large amounts of data efficiently, with 99% precision or higher.
Taking advantage of the best AI for the job, Piano generates data manipulation pipeline and analysis workflows that use the latest technologies. Importantly, Evidentli’s Agile AI uses generative AI to document and explain how these workflows are built so it is both transparent, and trusted.