Professor Roberta Wichmann, from the Master's program in Economics, published an article entitled "...in the form of..." along with researchers from USP and members of the IACOV-BR network. Improving the performance of machine learning algorithms for health outcomes predictions in multicentric cohorts in the journal Nature Scientific ReportsScientific Reports is the fifth most cited journal in the world, with over 696.000 citations in 2021, and receives widespread attention in policy papers and the media.https://www.nature.com/srep/).
The idea for this research arose when the authors wondered if the predictive performance of machine learning models could be improved simply by adding more data to the training in cases of predictions in the healthcare field. In other words, would it be possible to generalize a model to different hospitals in a state or different regions of Brazil? Thus, the authors tested different machine learning model training strategies, starting with local training using data from a single hospital, and progressing to different forms of aggregation with data from other hospitals, to evaluate the predictive performance in identifying the risk of death from COVID-19 and whether, therefore, it would be possible to generalize the prediction across different regions of the country.
How was the study conducted? Eight different strategies were created, and data from 18 hospitals across all regions of Brazil were used to train three machine learning models (xgboost, catboost, and lightgbm) in order to identify the best strategy to maximize predictive performance.
"In our study, the best strategy was training with data from a single hospital, achieving the best performance in 11 (61%) of the 18 hospitals, although in a few cases, better predictive performance was achieved by adding more data.", says the teacher RobertaTherefore, using data from a single hospital may result in better performance than adding data from different hospitals with different protocols and socioeconomic differences.
Congratulations Professor Roberta and all the authors and partners involved!
[vc_column_text][/vc_column][/vc_row][vc_row][vc_column][vc_btn title=”ACCESS THE ARTICLE HERE” color=”primary” link=”url:https%3A%2F%2Fwww.nature.com%2Farticles%2Fs41598-022-26467-6||target:%20_blank|”][/vc_column][/vc_row]