Month: November 2022

It may be counterintuitive to position AI as an accelerant to precision. And yet an under-exploited paradigm, Multimodal Learning combined with Composite AI, is powering discovery, resetting the benchmark for ‘real world’, and evidencing that precision medicine isn’t a pipe dream. If a broader range of stakeholders across Healthcare & Life Sciences are to benefit, our mainstream conversation needs to evolve. More SITS vendors must step up in support.

We know that a lot of work under the ‘AI’ banner has as many moving parts and shortfalls as it has roadblocks. We also know that embedded within the volumes of good data collected across HLS are swathes of rich but untapped information.

We have reached a tipping point: our current practice of training algorithms from a single modality (e.g. imaging, text, or stats) to validate a tightly defined query doesn’t do justice to our universal efforts to understand comorbidity, monitor disease recurrence, or advance personalised care, and drugs discovery.  

There is an alternative aligned approach, with successful field deployment:

• 𝐌𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠: Disconnected, heterogeneous (raw) data is harmonised, mined, and then unified into a single model;

• The backbone powering this is 𝐂𝐨𝐦𝐩𝐨𝐬𝐢𝐭𝐞 𝐀𝐈, which exploits a range of ML and non-ML techniques, in the context of the specific problem to be solved. This range works together sequentially to ‘fuse’ this data, by disassembling and normalising within one specific technique, before sending it on to another technique to apply further meaning – ‘passing the baton’, in a relay race, if you like. This multi-dimensional approach to tackling real-world problems delivers more granularity and scalability, which has eluded HLS.

Once synthesized, the insight is unified within a framework, which can be shared – several are available through open-source.

Blending The Known Unknowns with the Unknown Unknowns

The sheer range of data that can be exploited may seem over-whelming. However, multimodal isn’t a new approach, with pioneers across HLS already comfortably sharing their respective data with partners to advance.  

Equally, composite AI may seem daunting, but that need not be the case. The beauty of composite is its elasticity, in that only a few techniques need be applied for greater effect, in alignment to the range of data held, knowledge gaps that need to be filled, and respective skills base.

Acknowledged gains of Composite AI include:

• More granular information than RWE
• Deeper understanding at a molecular level
• Outperforms single modal AI
• Embeds Federal Learning and privacy-by-design
• Preservation of human knowledge within the loop ‘rounds out’ the data interpretation more deeply, thereby strengthening the analytical process
• Users/collaborations add ML/AI techniques at their own pace and capability

Explainability Versus Interpretability

We may all be familiar with the concept of Explainable AI (XAI), but it’s a generic reference that a whole range of stakeholders, each with differing priorities, are supposed to get their heads around.

This is where the notion of a trade-off in a pragmatic sense needs to be introduced to the mainstream conversation. As we apply some non-ML approaches techniques, which make up the other part of the Composite AI toolkit, in our quest for better detail, the black box comes fully into our front mirror. And we know the level of discomfort this provokes within the HLS community.

While some stakeholders may be comfortable with ML and automation, since these have been widely debated and evidenced, most will not be familiar of the inner workings of a composite model. While these HLS roles don’t need to be experts, there is a strong argument now for at least familiarising them with the basic principles, in relation to those techniques which lend themselves to high explainability, versus those that don’t.

For example, decision trees, which embed graphical representation, can convert data into a narrative of what’s going on. In contrast, Deep Learning, where the model trains itself to process and learn from data, yields low explainability, since the neural networks within are very difficult to interpret.

The trade-off here is whether to accept higher accuracy over single modal AI models, in exchange for less transparency behind the findings in complex data scenarios. We are after all still struggling to strike a balance between the flow of innovation, data fatigue, and actionable insights. This debate will continue for the foreseeable future.

With Multimodal Learning, No-One Acts Alone | Biotech Orchestrates; Pharma & Providers Collaborate

To relieve the burden of ML/AI maturity, this approach validates the need for Orchestrators from across HLS. The growing momentum behind Trusted Research Environments (TREs), systems thinking, and openEHR, across organisations and between research partners, is testament to this. These practices also help to position both Multimodal Learning and Composite AI in a constructive pragmatic light, as multi-disciplinary practices that advance great work among mutually respected peers.

SITS vendors have a role to play here too, alongside HLS teams, to champion the collection of good data, while working to ethics, non-bias, and responsibility. These are challenging but necessary codes of data management conduct that we must instil.

Current composite capability across HLS is already generating richer insight to supply novel intervention. Above are some of the key areas which stand to gain a lot, from healthcare systems such as the HSE in Ireland and NHS in Scotland currently rolling out Living Labs, through to the ongoing collaborative work across clinical trials and drugs discovery. With regulation in more European countries starting to acknowledge and reimburse digital therapeutics, this will, over time, pave the way to capture even more real world data at the edge.

One example is oncology, where multiple projects are exploiting multimodal data in one clinical area, and applying it to another, to better understand the correlation between some comorbidities – powerful stuff. Among these are the many Pharmaceutical leaders shifting their teams’ focus to targeted therapeutics designed for smaller patient cohorts.

Although in many, but not all instances, AI-first biotechs are the Orchestrators, they’re financially supported by a growing number of Pharma partners, via outcomes-based collaborations.

I’ve highlighted the biotech community in other posts as a cultural breath of fresh air. Yes, some parts of the sector are having a rough time at the minute, but there is some brilliant work coming through from pioneers, some of which feature above. The fact that they refuse to form exclusive relationships, yet are highly sensitive to IP preservation has been encouraging more from within the Pharma community to strike partnerships underpinned by Multimodal Learning and Composite AI.

From big tech, SaaS and vertical cloud players have been among the more recent additions to the Composite AI field, and all the signs, including investment, indicate they’re taking this seriously. More providers are joining. And this can only be good.

Composite AI isn’t a silver bullet. More SITS vendors can help to curate, steer, and widen their services offer. Skills support is also a solid way to build customer loyalty.

The Bottom line

We need a move away from the lean diet of ‘ML is doable in single mode; AI is elusive’.

We also need to put our burgeoning information pipelines to much netter use. Composite AI offers an effective path through which to advance at a pace and scale superior to most effort under way today. And while Composite AI isn’t for everyone, it can support multiple current and emerging use cases.

Further, as the concept of Ecosystem 2.0 evolves, Multimodal Learning and Composite AI will attract high calibre members, excited by the potential of what can be achieved together, with confidence.