By Deborah Borfitz
November 15, 2019 | A machine learning algorithm that contains the novel biomarker neutrophil gelatinase-associated lipocalin (NGAL) looks like a big win for patients with severe burns. It can help predict acute kidney injury (AKI) close to two days before clinical symptoms are detectable by conventional diagnostic measures alone.
That was the highlight of a co-presentation at the 2019 Next Generation Dx Summit by professor Hooman H. Rashidi, director of flow cytometry and immunology at the University of California Davis School of Medicine, and Nam Tran, associate clinical professor in the department of pathology and laboratory medicine at the university.
Rashidi kicked off the session with a brief overview on the application of machine learning in healthcare, which primarily involves supervised learning by techniques that include random forests, neural networks, nearest neighbors, support vector machines and, if the target is a number, logistic regression. The most accurate models achieve balance between over-fitting (low bias but high variance) and under-fitting (high bias but low variance), he says.
The data need to be vetted by experts and the proper features selected to build the models, Rashidi continues. Cross-validation of the platform is important in both reducing selection bias and avoiding over-fitting. Data is typically split 80/20 between training and validation sets.
Models also need to address a need. The M.D. Anderson Cancer Center/IBM Watson collaboration wasted $62 million by attempting to use machine learning to replace oncologists, who are already remarkably good at diagnosing cancer, rather than help primary care doctors in their role in cancer care, says Rashidi.
Critical Biomarker
So, what machine learning applications can be safely deployed in healthcare? The diagnosis of sepsis in severely burned patients is a strategic fit, according to Tran, given that up to 97% of those patients ultimately get the life-threatening condition. Early recognition is crucial.
The problem, says Tran, is that not all diagnostic criteria for sepsis apply to burn patients. Burn patients often develop systemic inflammatory response syndrome, which causes elevated body temperature, abnormally rapid breathing and changes in blood leucocyte count.
The lower hanging fruit turned out to be AKI detection in that same population. UC Davis researchers fed a machine learning algorithm data from 218 patients at four burn centers and identified 45% as having AKI, Tran says.
In a study published (DOI: 10.1016/j.jss.2015.03.033) in 2015, NGAL was found to be markedly increased in burn patients who develop AKI in the first week after injury and—unlike the usual diagnostic measures of urine output and creatinine level—is also an early independent predictor of AKI during acute resuscitation for severe burn injury. “When the kidneys are injured, they produce NGAL,” says Tran.
In a related study, just published in Burns, a simple nearest neighbor machine learning algorithm was shown to bump up the performance of urine output and creatinine for predicting AKI when NGAL is not available (NGAL has yet to be approved by the FDA for clinical use). NGAL and N-terminal B-type natriuretic peptide, another novel biomarker for AKI, were both used in the model, Tran says.
“NGAL is 90% accurate when combined with all other factors,” Tran says, and is the critical ingredient. Performance suffers if it’s not in the mix.
Other potential AKI biomarkers looked at, including kidney injury molecule 1 (KIM-1), have not worked well, says Tran.