medwireNews: Predictive models incorporating baseline variables can predict short-term tumor necrosis factor (TNF) inhibitor response with moderate to high accuracy in people with active ankylosing spondylitis (AS), study findings indicate.
Runsheng Wang (Columbia University Irving Medical Center, New York, USA) and colleagues believe their findings “may facilitate personalized treatment decision-making” in these patients.
The researchers used machine learning, based on data from 1207 individuals (mean age 39 years, 75% men) with active AS who received a TNF inhibitor while enrolled in one of six randomized clinical trials, to estimate the probability of each patient having a major response (ASDAS ≥2.0) or no response (ASDAS <1.1) to treatment at 12 weeks.
They initially included 21 variables in five models, but then reduced this to two models – logistic regression and random forest – with fewer variables for practical reasons.
The logistic regression model included C-reactive protein level, patient global assessment (PGA), BMI, BASDAI question 2 score (ie, severity of neck, back, and hip pain), and BASFI score, while the random forest model included CRP, BMI, and BASDAI question 2 score.
The models accurately predicted a major response, which occurred in 33.7% of patients, 72–74% of the time, and accurately predicted no response, which occurred in 34.3% of patients, 74–75% of the time. The models had good specificity (85–89%) and moderate sensitivity (45–46%) for a major response, as well as good specificity (90%) and moderate sensitivity (44–45%) for no response.
Wang et al report in JAMA Network Open that the likelihood of a major response increased with increasing CRP level, PGA score, and BASDAI question 2 score, and decreased with increased BMI and BASFI score.
For no response, the probability increased with increasing age and BASFI score, whereas increased CRP level, PGA score, and BASDAI question 2 score were associated with a decreased likelihood of no response.
The investigators observed similar results when they validated their models in an additional 692 participants (mean age 38 years, 77.0% men) from four studies.
They therefore conclude that the models may enhance confidence in choosing TNF inhibitor therapy in individuals with a high probability score for major response.
They add: “Absence of a response in a patient predicted to have a high probability may raise a question about adherence to treatment.
“Conversely, a course of TNF [inhibitor] treatment may be terminated quickly if nonresponse occurs in a patient predicted to have a high probability of no response.”
medwireNews is an independent medical news service provided by Springer Healthcare Ltd. © 2022 Springer Healthcare Ltd, part of the Springer Nature Group
JAMA Netw Open 2022; 5: e 222312