A machine-learning algorithm can use demographic, symptomatic, and clinical data to accurately predict the health-related quality of life (QOL) of patients with kidney stones, new research shows.
“This can help us determine whether or not to do surgery on a patient when their kidney stone is not too painful,” said Naeem Bhojani, MD, from the University of Montreal, Canada.
The algorithm his team developed uses clinical data from a patient’s electronic health record to arrive at a score, eliminating the need for a questionnaire.
“If you ask urologists what is most important when it comes to kidney stone disease, they’ll say it’s the removal of the stone; we want the stone gone,” Bhojani told Medscape Medical News. “But that doesn’t necessarily equate with a better quality of life.”
Some patients are asymptomatic and can carry a stone without it bothering them. “Their quality of life is excellent, so why do surgery?” They can be followed with observation, he said.
However, a lot of variables are involved in an assessment of kidney stones. “It’s a very intricate measure,” Bhojani explained.
In fact, it was the “horrible quality-of-life scores” derived from standardized questionnaires completed by patients with kidney stones that led nutritionist Kristina Penniston, PhD, from the University of Wisconsin-Madison, and her team to develop a tool to capture the unique symptoms and challenges of urolithiasis.
“At the time, there were no disease-specific instruments to assess these patients,” she said. There are hundreds of quality-of-life tools for many different diseases, and a lot of them “are not as painful or destructive as kidney stones,” she explained. And about 10% of the population is affected by kidney stones, which is about the same proportion of the population affected by diabetes.
Today, the Wisconsin Stone Quality of Life Questionnaire (WISQOL) “is being widely used,” Penniston said. Researchers are gathering data, tracking changes in stressors and the presence or absence of stones, and documenting reactions and changes during disease progression.
“It’s an episodic disease,” Penniston said. “It causes disruptions to family life and work life, and it brings on anxiety. Some patients feel disabled waiting for the next stone.”
Some patients worry about stones all the time, whereas others have no issues. But there “are a lot of people in between.” Knowing how they are managing helps surgeons make decisions about what to do clinically, she said.
The WISQOL was long overdue. “What’s special about stone patients is that they have lower quality of life than someone without stone disease, even when they are asymptomatic,” said Bhojani.
There are many factors to explain that. “The idea that you have a stone in your kidney — knowing how painful it can be, even if you’re asymptomatic — is a ticking time bomb,” he said.
And patients with kidney stones are in and out of the hospital and have to follow a strict diet. “These factors alone affect quality of life,” he noted.
Bhojani’s Montreal-based team is a member of the consortium validating the WISQOL and they offered to translate the questionnaire into French. But as the researchers got involved, they started to think about how machine learning could use the data that had been collected to predict quality-of-life scores for stone patients.
Encouraging Results
The team used 3-, 12-, and 24-month data collected from 3206 patients with kidney stones, ages 18 years and older, who completed the 28-item WISQOL questionnaire at 16 outpatient centers in Canada and the United States between June 2014 and May 2016. Those patients also provided demographic and medical information, including surgical status, age at first onset, body mass index (BMI), number of stones, family history, use of potassium citrate, and history of depression or anxiety.
Using gradient-boosting and deep-learning models, the researchers allotted 70% of the data for training, 10% for validation, and 20% for testing, which is the machine-learning standard.
The primary outcomes were health-related quality-of-life-score and quintile estimates, which were derived from Pearson’s correlation regression performance and area under the receiver operating characteristics curve (AUC) classification performance.
With clinical data only, the deep-learning toolset obtained a correlation of 0.59 and an average AUC of 0.70 for the five quintiles.
“The model was able to distinguish between lowest and highest quintile and correctly weighted symptomatic status, BMI, age, as well as other medical and demographic features, to estimate QOL,” David-Dan Nguyen, a medical student at McGill University in Montreal, said this week during his presentation of the study data at the virtual European Association of Urology (EAU) 2020 Congress.
Five factors — BMI, age, age at first onset, symptomatic status, and emergency-department visits in the previous 4 weeks — weighed more heavily in the predictions. The model performs best when identifying extremes; the AUC was 0.79 for the lowest quintile and 0.83 for the highest, Nguyen explained. “It underperforms when it comes to classifying patients in the middle quintile of health-related quality of life.”
That’s something that the researchers will continue to work on. As more data become available, the machine-learning algorithm will become more precise.
During a live video session at the EAU conference, chair Patrick Krombach, MD, from University Hospital Mannheim in Germany, asked Nguyen whether this could be “used to test different kind of strategies, if you have a patient with a stone and you want to check out the best one?”
“In the future, with more data coming from the Wisconsin quality-of-life study, hopefully our model can improve and be used in more refined clinical settings,” Nguyen explained.
The tool could also help urologists who don’t have the resources to offer questionnaires, and will save consultation time, allowing physicians to see more patients, Bhojani added.
Bhojani, Penniston, Nguyen, and Krombach have disclosed no relevant financial relationships.
European Association of Urology (EAU) 2020 Congress: Abstract 785. Presented July 26, 2020.
For more news, follow Medscape on Facebook, Twitter, Instagram, and YouTube.