What is the clinical evidence from this study?

Study design: Observational. Population: Lupus nephritis (n=173). Intervention: Machine learning prediction using clinical indices vs. Renal biopsy (actual pathology). Primary outcome: Total classification accuracy for three-class lupus nephritis pathology using random forest.

July 2, 2018Open Access

Lupus nephritis pathology prediction with clinical indices

YTYouzhou TangCentral South University WZWeiru ZhangCentral South University MZMinfeng ZhuUniversity of Technology Sydney

Key Result

A machine learning model using clinical indices predicted the pathological classification of lupus nephritis with 51.3% total accuracy across three classes, and successfully predicted acute and chronic indices.

Study Design

Type

Observational (n=173)

Multicenter

Yes

Structured PICO

Can machine learning models using clinical indices accurately predict lupus nephritis pathological classification and acute/chronic indices?

Population

173 patients (mean age 27.5 years, 84.7% female) with lupus nephritis who underwent renal biopsy, whose clinical and laboratory indices were retrospectively analyzed to train machine learning models.

Exposure

Machine learning models (random forest for classification, multiple linear regression for acute and chronic indices) using clinical and laboratory indices.

Comparator

Renal biopsy (gold standard) histological classification and acute/chronic indices.

Outcome

Accuracy of pathological classification (class II, III&IV, V) and prediction of acute index (AI) and chronic index (CI).surrogate

Machine learning models using clinical indices show potential as a noninvasive auxiliary tool for predicting lupus nephritis pathology and acute/chronic indices, though accuracy requires further improvement.

Limitations

Prediction accuracy instability, with less satisfactory efficiency for class V
Uneven sample distribution across pathological classes
Cannot replace renal biopsy as the gold standard
Retrospective data collection

Abstract

Abstract Effective treatment of lupus nephritis and assessment of patient prognosis depend on accurate pathological classification and careful use of acute and chronic pathological indices. Renal biopsy can provide most reliable predicting power. However, clinicians still need auxiliary tools under certain circumstances. Comprehensive statistical analysis of clinical indices may be an effective support and supplementation for biopsy. In this study, 173 patients with lupus nephritis were classified based on histology and scored on acute and chronic indices. These results were compared against machine learning predictions involving multilinear regression and random forest analysis. For three class random forest analysis, total classification accuracy was 51.3% (class II 53.7%, class III class III class V 61%. Additionally, machine learning selected out corresponding important variables for each class prediction. Multiple linear regression predicted the index of chronic pathology (CI) (Q 2 = 0.746, R 2 = 0.771) and the acute index (AI) (Q 2 = 0.516, R 2 = 0.576), and each variable’s importance was calculated in AI and CI models. Evaluation of lupus nephritis by machine learning showed potential for assessment of lupus nephritis.

Demander à l'IA

Bookmark

View Full Paper