How machine learning can forewarn about disasters


In his iconic best-seller  The Tipping Point, American author Malcolm Gladwell talks about how a bunch of kids wore Hush Puppies shoes, then an almost-dead brand, and triggered a wave of sales that made the shoes a rage across the US.

A tipping point is when a system changes its state suddenly and irretrievably. For long, no changes in input parameters (say, temperature) seemed to cause any impact on the system — a phase known as ‘scale invariance’ — but, abruptly, ‘scale invariance’ ceases to exist, and the system has changed for good.

So, the tipping point is much like the ‘last straw that breaks the camel’s back’ — weight kept building up on the animal’s back, until a lightweight straw proved enough to finally crack it.

The principle of ‘tipping point’ — called ‘critical transitions in complex systems’ — is sought to be used in predicting many things in life such as onset of diseases, stock market crashes, and climate change. It is still work-in-progress, but the world is definitely cruising towards developing early warning systems.

For example, Prof RI Sujith of the Department of Aerospace Engineering, IIT-Madras, has used this principle to build a model that can predict an impending blowout in aircraft and rocket engines, “enabling us to take action in time to evade it”.

To be able to time a tipping point, it may be good to divine its precursors. For example, events at the cellular or molecular level — such as gene expression or protein synthesis — slow down before the tipping point. Scientists discern a link (in some cases) between protein synthesis and bacterial growth.

Such ‘critical slowing down’ (CSD) is a focus area of research. In an April 2022 paper titled ‘Identifying critical transitions in complex diseases’, published in the  Journal of Biosciences, Smita Deb, et al, of the Indian Institute of Technology, Ropar, note that “CSD is the phenomenon where the system’s return to the current equilibrium state is slowed down upon perturbations in the vicinity of a tipping point”. One of the authors of the paper, Prof Partha Sharathi Dutta of the Department of Mathematics, told  Quantum, “Now we know fairly well that CSD-based early warning signals can forecast an upcoming tipping point.” However, he cautioned that more research is needed to determine precisely “when” the transition will occur.

Powered by machine learning

Dutta said that while the concept of ‘tipping point’ has been around for decades, the advent of machine learning has rejuvenated it for research in developing early warning systems.

In their paper, the authors say they have developed “a novel detection method, using simple theoretical models to train a deep neural network to detect critical transitions — the Early Warning Signal Network (EWSNet)”.

This network, trained on simulated data, can reliably predict observed real-world transitions in systems ranging from rapid climatic change to the collapse of ecological populations. “Our work highlights the practicality of deep learning for addressing further questions pertaining to ecosystem collapse, and has much broader management implications.”

In essence, the ‘critical transitions in complex systems’ is an old idea now energised by ML, which holds a lot of promise. Prof Sujith points out that the ‘complex system theory’ has been well adopted by the medical research community.

However, it still appears to be in the ‘research’ stage.

Need for data

Yet another author of the paper, Dr Mohit Kumar Jolly, Assistant Professor at the Centre for BioSystems Science and Engineering (BSSE), Indian Institute of Science, Bengaluru, said that in biomedical science most research is now retrospective — finding out the precursors to a tipping point.

Jolly is working on ‘metastasis’ (spreading of cancer). “We would ideally like to look at the data of a patient from his primary tumour and predict whether he would develop metastasis or not,” he said, “but we have not reached that stage yet.” He, however, noted that proof-of-concept has been established and “we are cruising towards that stage”.

Dr Dutta is more circumspect, though he too certifies that the field shows a lot of promise with the advent of machine learning. Nonetheless, the data required for training the machine is huge — “if the machine is trained well, it can forecast well”. However, that will take a few more years, he said.