Machine learning may put an end to multiple doctor visits and endless medical tests

Medicine has opened new fields of research where machine learning could be applied, doctors and data scientists have teamed up to develop new tools for practical applications in hospital settings. From the use of virtual reality, to the crunching of numbers, machine learning scientists have found medicine is a treasure-trove full of data for the taking. A lot of that work has concentrated in the research for tools that can help doctors do an estimated prognosis on a patient. To be clear, machine learning is not trying to replace doctors, instead it is trying to take the technology to a mature stage where it can be applied in medicine.

It may be difficult for doctors to develop a prognosis with little information, and predicting the future is not a luxury they can afford. They can recommend a course of action and monitor the patient’s reaction then change or tweak the treatment to enhance the patient’s prognosis, assess again, then repeat if no improvements are observed. This means patients will have to come back for more tests, consultations, prescriptions, and so on. Unfortunately, there is no other way, even when we know the stages of progress of diseases, it is not possible to predict how a particular patient will react to the planned treatment. Could there be a tool doctors could use to at least have a likely scenario? With a probability attached to it, maybe?

This is where machine learning enters the picture, but if doctors want to use machine learning to make good estimates then they will need more adequate data. This is exactly what many researchers are doing, for example consider a recently published study conducted by several investigators affiliated to hospitals across the country (Marshall University Joan C. Edwards School of Medicine, The Cleveland Clinic Foundation, John’s Hopkins University, and New York Medical College), they used machine learning to predict progressive kidney failure in several patients that were already registered in the “Modification of Diet in Renal Diseases” or MDRD study previously conducted by other researchers. They used different machine learning methods (generalized linear model, support vector machine, decision tree, feed-forward neural network and random forest evaluated within the context of 10 fold validation), they conducted the analysis using the CARET package available within R software.

Their primary goals were the prediction of different outcomes: the composite outcome of death, dialysis or doubling of serum creatinine using the modification of diet in renal disease. The data set had 25,903 records but they went through a process of patient selection and they finally used 692 records only. The data was refined and then analyzed using the R software. They concluded that all the models studied succeeded in predicting prognosis, to their surprise each of the models showed that specificity was superior to sensitivity. Accuracy ranged between 66 and 77%. One important result from this study was the determination of the most influential variable in the data. The baseline serum creatinine was among the three basic variables in all of the models, it showed at the top in the GLM, the SVM and the RFor models. They were not surprised by this results since the initial renal function was expected to predict outcomes in this population with chronic kidney disease. The results also showed that dietary protein and blood pressure did not achieve great importance in these different models, which did not surprised the research group since these interventions did not significantly affect outcomes. The great outcome from this analysis was that all the models they used performed satisfactorily.

These research results are very promising but more data and research are needed before machine learning based technologies are extensively used in medicine in a routine basis. As datasets become available and more researchers are willing to seize this knowledge, the technology development will just keep moving forward. This could be of great help to both patients and physicians.


Reference

Predicting adverse outcomes in chronic kidney disease using machine learning methods: data from the modification of diet in renal disease. Zeid Khitan, Anna P. Shapiro, Preeya T. Shah, Juan R. Sanabria, Prasanna Santhanam, Komal Sodhi, Nader G. Abraham, and Joseph I. Shapiro. Marshall Journal of Medicine, Volume 3, Issue 4, Article 10, 2017.