With the increasing use of artificial intelligence (AI) in radiology, it is essential to minimize bias in machine learning systems before implementing their use in real clinical scenarios. This is suggested by a special report published in the journal Radiology: Artificial Intelligence which targets ineffective practices relating to the collection and management of data.
And major report published in the Journal Radiology: artificial intelligence aims to identify and minimize the biases that may be found in machine learning models before they are put into production under clinical conditions. This is the first in a series of three reports, which describes suboptimal practices used in the data processing phase of machine learning system development and presents strategies to mitigate them.
Machine learning models biased by often ineffective data management
” There are 12 suboptimal practices that occur during the data processing phase of developing a machine learning system, each of which can predispose the system to biases.announces Pr Bradley J. Erickson, professor of radiology and director of the AI laboratory at the Mayo Clinic, in Rochester -Minnesota, USA -. If these systematic biases are not recognized or accurately quantified, suboptimal results will ensue, limiting the application of AI to real-world scenarios. »
Prof. Erickson adds as a preamble that the subject of proper data management has been attracting more and more attention, but that guidelines on the proper management of big data are scarce. ” Regulatory challenges and translational gaps still hinder the implementation of machine learning in real clinical scenarios, he continues. However, we expect the exponential growth of AI systems in radiology to accelerate the removal of these barriers. To prepare machine learning systems for adoption and clinical implementation, it is essential to minimize bias. »
An expert report that targets 12 practices to be improved to validate AI models
In this report, Prof. Erickson and his team suggest mitigation strategies for the 12 suboptimal practices that occur in all four data processing stages of AI system development – three for each data processing stage. -.
Regarding data collection, incorrect identification of data set, unique data source and untrusted data source are targeted. For data mining, researchers want to work on analyzing inadequate exploratory data, exploratory data without domain expertise, or failure to observe the actual data.
To improve data splitting, the report identifies leakage between datasets, unrepresentative datasets, as well as hyperparameter overfitting. Finally, on the topic of data engineering, researchers focus on the incorrect deletion of features, their incorrect resizing or poor management of missing data.
Careful collection planning and collaboration with data science experts
Professor Erickson also finds that medical data is often far from ideal for feeding machine learning algorithms. ” Each of these steps could be subject to systematic or random biaseshe adds. It is the responsibility of developers to accurately manage data in difficult scenarios such as data sampling, anonymization, annotation, labeling and handling of missing values. »
The report recommends careful planning before data collection including a thorough review of clinical and technical literature, as well as collaboration with data science expertise. ” LMultidisciplinary machine learning teams should have members or leaders with both data science and clinical domain expertise he says.
To develop a more heterogeneous training dataset, Prof. Erickson and his co-authors suggest collecting data from multiple institutions in different geographic locations, from different vendors and from different time periods, or including public datasets. .
Two other reports that will target the development and evaluation phases of AI models
” Building a robust machine learning system requires researchers to do some detective work and look for ways the data can fool you.he declares. Before putting data into the training module, you should analyze it to make sure it reflects your target population. The AI won’t do it for you. Even after excellent data management, AI systems can still be subject to significant biases. The second and third reports in this series published in Radiology will focus on biases that occur in the model development and evaluation, as well as reporting phases.
In recent years, machine learning has demonstrated its usefulness in many areas of clinical research, from image reconstruction to improving diagnostic, prognostic and monitoring tools, concludes Professor Erickson. This series of reports aims to identify erroneous practices in the development of machine learning and to mitigate them as much as possible. »
Bruno Benque with RSNA