The high diagnostic and prognostic application values of newer IgAN machine learning models suggest their eventual utility in clinical practice.
Credit: Fotolia
Findings from a recent systematic review and meta-analysis are providing valuable insight into the diagnostic and prognostic application values of new IgA nephropathy (IgAN) machine learning models.1
The study, coined by investigators as the first meta-analysis to explore the application of machine learning for IgAN diagnosis, prognosis, and IIgAN-RPT validation, found newer IgAN models generally outperformed traditional ones.1
“Unveiling the predictive performance of machine learning in IgAN can help guide personalized treatment, determine follow-up frequency, and minimize unnecessary use of immunosuppressants in low-risk patients,” Kaiting Zhuang, of the department of nephrology at the First Medical Center of Chinese PLA General Hospital, and colleagues wrote.1 “Accurate prediction tools are especially important for heterogeneous diseases like IgAN.”
A progressive kidney disease and the most common primary glomerulonephritis, IgAN’s clinical course varies from person to person, with some patients not experiencing symptoms for years on end while others more rapidly progress to kidney failure. Prompt diagnosis and treatment are essential for improving outcomes, and research has shown machine learning may be a viable diagnostic and prognostic tool in IgAN.2
To analyze the use of machine learning for the diagnosis and prognosis of IgAN, investigators searched Embase, Pubmed, Cochrane Library, and Web of Science until February 24, 2024, for publications on machine learning-based diagnosis and prognosis of IgAN. For inclusion, studies were required to meet the following criteria:
From each eligible study, investigators extracted the total number of samples and the number of events in the training and validation sets, the C-index and their 95% confidence intervals (95% CI) or standard errors, sensitivity, specificity, accuracy, calibration slope, net reclassification index (NRI), and integrated discrimination improvement.1
Acknowledging differences in the variables included in machine learning models and the inconsistent parameters, investigators deemed a random-effects model to be preferable for the meta-analysis, and subgroup analyses were performed according to model type, variable type, endpoint definition, and follow-up time.1
A total of 47 studies involving 51,935 patients were included in the analysis, 12 of which constructed diagnostic models and 27 of which constructed prognostic models. Additionally, a single study constructed models for both diagnosis and prognosis, and another 7 studies conducted external validation of the 2019 IIgAN-RPT tool. The meta-analysis of IgAN diagnosis, IgAN prognosis, and IIgAN-RPT validation involved 7270, 36,659, and 8006 patients with a total of 38, 162, and 19 models, respectively.1
Among the 38 diagnostic models, 27 provided a C-index, and 15 provided sensitivity and specificity. The pooled C-index of the 27 diagnostic models was 0.902 (95% CI, 0.878–0.926) in the training set and 0.851 (95% CI, 0.808–0.894) in the validation set. The overall sensitivity and specificity were 0.82 (95% CI, 0.78–0.86) and 0.81 (95% CI, 0.71–0.88) in the training set, and 0.82 (95% CI, 0.78–0.86) and 0.81 (95% CI, 0.71–0.88) in the validation set, respectively.1
Among the 162 prognostic models, the pooled C-index for model discrimination of 144 prognostic models was 0.838 (95% CI, 0.827–0.850) in the training cohort and 0.817 (95% CI, 0.801–0.833) in the validation cohort. The overall sensitivity and specificity of models that provide sensitivity and specificity were 0.81 (95% CI, 0.76–0.85) and 0.87 (95% CI, 0.83–0.90) in the training set, and 0.88 (95% CI, 0.78–0.93) and 0.88 (95% CI, 0.82–0.92) in the validation set.1
The 87 survival models (COX regression) had a C-index of 0.826 (95% CI, 0.815–0.837) in the training set and 0.828 (95% CI, 0.810–0.845) in the validation set, indicating survival models had favorable discriminative ability. In the rank sum test for non-survival models, results showed the logistic regression model (C-index, 0.840; 95% CI, 0.785–0.989) did not outperform other machine learning methods, except the Naïve Bayesian model (C index, 0.653; 95% CI, 0.543–0.763; P <.05).1
The meta-analysis of the external validation of IIgAN-RPT (19 models) showed a pooled C-index of 0.801 (95% CI, 0.784–0.817). Of note, 3 studies verified the model's performance improved after race was included in the modeling variables, and their NRI was 0.52 (95% CI, 0.33–0.72), 0.13 (95% CI, 0.08–0.29), and 0.49 (95% CI, 0.41–0.59), respectively.1
Investigators acknowledged the lack of clarity regarding the confidence interval of the C-index of some models and the high risk of bias in the included studies as potential limitations to these findings.1
“Machine learning can help physicians to diagnose IgAN and assess the subsequent prognosis,” investigators concluded.1 “Expanding the application of three-dimensional reconstruction techniques in diagnostic models, using albuminuria as a more sensitive prognostic endpoint, enhancing moderate-risk prognosis, extending the racial validation scope and pediatric validation of international tools, and translating the model into clinical calculators will be the future direction.”
References