Arthritis & Rheumatism, Volume 62,
November 2010 Abstract Supplement

Abstracts of the American College of
Rheumatology/Association of Rheumatology Health Professionals
Annual Scientific Meeting
Atlanta, Georgia November 6-11, 2010.


Machine Learning Models Using Multiple Low Abundance Protein Biomarker Levels Are Superior to Those Using Clinical Laboratory Values in Diagnosing ISN/RPS Class of Lupus Nephritis.

Oates2,  Jim C., Petri1,  Michelle A., Almeida4,  Jonas S., Fleury3,  Thomas W., Michael3,  Janech G., Arthur3,  John M.

Timonium, MD
Medical University of South Carolina, Charleston, SC
Medical University of South Carolina
The University of Texas M. D. Anderson Cancer Center

Objectives:

Treatment in lupus nephritis (LN) is often driven by renal biopsy findings, and traditional biomarkers are not predictive of renal pathology. We hypothesized that levels of multiple candidate low abundance urine proteins, when analyzed my multivariable machine learning techniques, would create more effective models of International Society of Nephrology/Renal Pathology Society class of nephritis (ISN/RPS Class) than traditional biomarkers now available to clinicians.

Methods:

Subjects from the Charleston and Baltimore LN inception cohorts and the Rituxan in LN (LUNAR) study population were recruited. ISN/RPS Class was determined prior to induction therapy. Urine samples were collected at entry for analysis. Urine levels of 12 candidate low abundance proteins (chemokines, growth factors, cytokines, and renal damage markers) were determined by the multiplex bead array, ELISA or activity assay for all patients. Levels of individual markers were used to create multivariable models of LN Class at baseline. Diagnostic models were trained using machine learning (artificial neural network (ANN) and nearest related neighbor (NRN) algorithms). The diagnostic power of models was reported as the receiver operating characteristics curve area under the curve (ROC AUC), with AUC values for perfect and non-diagnostic tests being 1 and 0.5 respectively. Input variables were traditional biomarkers (anti-double stranded DNA antibodies (DNA), C3, C4, serum Cr) and/or the selected biomarker panel. The output variables were the individual biopsy classes.

Input variablesNumber of subjectsOutcome (ISN/RPS Class)NRN ROC AUCANN ROC AUC
Novel biomarkers99II0.910.95
 III0.820.68 
 IV750.92 
 Proliferative0.840.94 
 V0.590.94 
Traditional Biomarkers66II0.820.97
 III0.640.78 
 IV0.620.68 
 Proliferative0.630.96 
 V0.640.57 
Novel and traditional biomarkers66II0.840.82
 III0.790.85 
 IV0.820.96 
 Proliferative0.730.89 
 V0.860.74 

Results:

Biomarker data were available for 99 subjects. Combined biomarker clinical data were available for a subset of 66 subjects. In general, ANN and NRN models using biomarkers either alone or in combination with traditional biomarkers were superior to models with clinical variables alone. ANN models tended to have greater diagnostic power than NRN models. 93% of the predictive power from the ANN model of proliferative disease came from GM-CSF, IFNa2 MCP1, NGAL, IL-6, and IL-12 levels.

Conclusions:

This study suggests that multiple biomarkers representing diverse pathogenic mechanisms by machine learning modeling techniques are effective in diagnosing ISN/RPS class of nephritis. It demonstrates that when markers of multiple types of cell activation, migration, and damage are combined into a single model, diagnostic power is improved over models using traditional biomarkers alone.

To cite this abstract, please use the following information:
Oates, Jim C., Petri, Michelle A., Almeida, Jonas S., Fleury, Thomas W., Michael, Janech G., Arthur, John M.; Machine Learning Models Using Multiple Low Abundance Protein Biomarker Levels Are Superior to Those Using Clinical Laboratory Values in Diagnosing ISN/RPS Class of Lupus Nephritis. [abstract]. Arthritis Rheum 2010;62 Suppl 10 :1403
DOI: 10.1002/art.29169

Abstract Supplement

Meeting Menu

2010 ACR/ARHP