Machine Learning in Healthcare: Humans Need Apply
Machine learning (ML) has come a long way since its first application in the 1955 Samuel Checkers-playing Program. It has seemingly touched all aspects of our lives, our health included. ML has made healthcare more efficient and more reliable in many ways. Innovative machine learning tools can quickly and accurately help assess patient health and aid physicians in clinical care decision making.
ML algorithms, a subset of artificial intelligence, are designed to make predictions or decisions based on patterns in datasets. Today, we have access to more patient data and biological data than ever before (often called "big data"). ML algorithms are built to extract meaningful information by detecting patterns or features in big data. One example of this is in cancer pathology where machine learning tools can identify tumor cells more accurately and with more concordance than the human eye in a tissue biopsy (Beck et al., 2019). However, that does not mean humans are not needed.
To perform at its best (or perform at all), ML needs human ingenuity, input, and constant evaluation. To create a ML algorithm, a data scientist (a human) must first write a computer code. The code provides the instructions for the algorithm and is the basis for its ability to solve problems or predict outcomes.
To get started data scientists require “training data” . Training data consists of already-known outcomes and information relevant to those outcomes. Data scientists will first tweak the code and algorithm to have high accuracy on the training data where they know the answer. After which, the algorithm is applied to "test" data where outcomes are unknown. Predictions on the test data are based on patterns detected in the training data.
Thus, thoughtfully curating the training data is a critical step in developing ML tools that can make reliable predictions and decisions. This calls for collaboration between data scientists, researchers and physicians to ensure the algorithms are trained on the highest quality data. In sum, good data in, good predictions out (or the reciprocal, garbage in, garbage out).
Once ML algorithms are trained and have demonstrated high predictive accuracy, that still doesn’t mean algorithms should have total authority. Just as pilots monitor flights with autopilot capabilities and can intervene if necessary, medical professionals must monitor and evaluate the outcomes. Medical professionals knowing and trusting ML data will be critical to improve the experience and outcome of care.
At Olaris, we develop ML algorithms that can predict which drug will work for a specific individual. Those algorithms are built by careful collaboration with data scientists, metabolite scientists and medical professionals. We believe only through this highly iterative and united approach can we identify signatures that our stakeholders can trust to impact patient care.
ML algorithms are human-driven from beginning to end. From the code they are based on and the data they are trained on, to the refinement of their predictions and interpretation of the results -- all require human inputs. ML will continue to transform and hopefully improve how we live our lives. To make that a reality -- Humans Need Apply!
Beck A., et al., An empirical framework for validating artificial intelligence-derived PD-L1 positivity predictions using samples from patients with urothelial carcinoma [abstract]. In: SITC, 2019 Nov 6-10; Maryland. Boston: PATHAI; 2019. P730.