Open The Podbay Doors, HAL. Will Health Care Will Struggle To Adopt AI?

Remember in Space Odyssey, 2001, when HAL started to question Dave and the other crew members and ended up refusing to let him back on the…

Open The Podbay Doors, HAL. Will Health Care Will Struggle To Adopt AI?

Remember in Space Odyssey, 2001, when HAL started to question Dave and the other crew members and ended up refusing to let him back on the spaceship? Well, in AI, we are possibly getting to this point where AI is making judgements which go against our reasoning. But, who is correct?

Health care innovation

Having been involved in healthcare research for more than two decades, there’s one thing I know … few sectors are more resistant to technological change than healthcare. In fact, there was a quote at one time that the magic number was seven in health care innovation, as that was the number of years that it took to show that a flagship wave of technological innovation in the NHS had failed and then be replaced with another flagship policy from the government.

The waste of money in Connecting for Health [here] showcased how resistant the NHS is to technological change. It ran from 2003 to 2015 and cost over £20bn — and with little in return. And, so, while we see successful pilots of AI in the NHS, one must wonder if they will ever properly see the light of day at scale in our world of health care.

This resistance is highlighted in a new paper by Agarwal et al. outlines an experiment in using AI within diagnoses [1]:

The core finding is that AI was actually more accurate than two-thirds of radiologists in the diagnoses, but when radiologists used AI to help them, they found that they often ignored the advice when it conflicted with their own.

The following table shows the summary results from a number of radiologists related to the ground truth and the deviation from ground truth, along with the deviation from the AI decision.

Figure 1 [1]

A variation away from ground turn leads to incorrect decisions and made correct recommendations for 70% of the observation and for an average decision time of around 2.8 minutes. The average deviation from ground truth ranged from 0.191 to 0.232.

Figure 1 outlines the distribution of accuracy for radiologist assessments compared to AI’s accuracy. RMSE is the root mean squared error and AUROC is the Area Under the Curve of the Receiver Operating Characteristic. The average AI measurement is illustrated with the dotted line in the distribution.

Figure 2 [1]

Conclusions

There are some things in our roles that machines are good at spotting and where they can gather lots of evidence to prove or disprove something. For almost the first time, machines are now improving on expert advice, and we must thus decide whether we want to leave some tasks to them or whether they are just another tool in the armoury. For health care, the resistance to technology is ever-present. I have no online healthcare record and very little in the way of a digital footprint related to my health. In fact, the only paper forms I have had to complete in the two few years are related to the NHS.

The lack of investment in digital technologies in the NHS to a symptom of many things, including the lack of vision, a lack of proper integration across the NHS, and the siloing of healthcare systems. But it is also a factor of human resistance to the usage of technology-driven methods.

Reference

[1] Agarwal, N., Moehring, A., Rajpurkar, P., & Salz, T. (2023). Combining Human Expertise with Artificial Intelligence: Experimental Evidence from Radiology (No. w31422). National Bureau of Economic Research.