"Machine learning" and clinical practice

@varelalaf

Machine learning is a branch of artificial intelligence based on the idea that systems can learn, identify patterns and make decisions with minimal human intervention, as is the case, for example, with weather forecasting. When I was little, the men of the time (then there were no women in these jobs) drew isobar lines on a blackboard where they had made their predictions based only on their experience and the four sources of data that came to them from the meteorological observatories. Now, as the forecasts are made by the models, meteorologists have enough time to comment on photos sent by viewers, in addition to raising awareness of the effects of pollution and climate change.

As is known, machines don’t have common sense. Therefore, they need much more data than humans to learn a task. The difference is that when the models have developed an algorithm, based on swallowing mountains of data, they are able to propose very effective solutions, since intuition does not play tricks on them, and they are also not susceptible to making mistakes.

In a context of machine learning, advancing through all areas of human activity, a team from Harvard Medical School has analyzed to what extent the phenomenon is being introduced (or will be) into clinical practice, an analysis of which I have made into an extract:

The Diagnostic Process. The models could support a differential diagnosis by suggesting possibilities that are beyond the knowledge of a specific doctor, especially in the most complex, and strangest cases, in which diagnostic errors are frequent.

The Supplementary Tests. This is a very active field of advance for machine learning, where quite a few innovations are already being seen, such as analysis of radiological images, ECG, tissue analysis, cytology, back of the eye, remote dermatology or colon polyps.

Forecasting. One of the most cumbersome clinical activities for doctors is having to predict the evolution of a certain clinical process, since statistics are one thing and it is quite another to guess what will happen to a specific patient, a situation that, adds up to an environment of expectation and anxiety. For this reason, having personalized forecasts prepared from large databases is of great help.

Choice of Treatment. For certain diseases, offering patients the most suitable treatment requires prescribing doctors to be highly up-to-date and able to consistently explain the balance between effectiveness and adverse effects for each of the options. Support for machine learning in this field is very useful.

Clinical Research. Machine learning becomes a very practical instrument in offering a helping hand in the search for effectiveness in the real world, in the sense that, once the criteria for patient recruitment are defined, doctors receive notices at the precise moment that they are serving people with the profile sought by researchers.

Limitations of machine learning for clinical practice

Machine learning is here to stay, and therefore doctors and nurses should see this trend as a chance to free themselves from a lot of bureaucratic and technical tasks of little interest, to focus more on patients and to raise the quality of their work. The Harvard Medical School report states, however, that the main limitation for the introduction of machine learning in healthcare workers is the poor quality of the data from not only fragmented medical records, but often show deficiencies in the accuracy of the information and, what is worse, in the fidelity of what is collected.

Imagine that weather model had data from observatories that measured biased temperatures, leaking rain gauges, and rusty anemometers. Then,  weather men and women could dedicate a lot of effort to answering claims, instead of commenting on the clouds of the beautiful photos that viewers send, now entirely involved in the meteorology show.

Jordi Varela
Editor