Using Machine Learning to Triage Bone Marrow Specimens to Improve Performance of Plasma Cell Disorder FISH Testing. (21st September 2018)

• Record Type: Journal Article
• Authors: Linder, Grace; Sohani, Aliyah; Baron, Jason
• Is Part Of: American journal of clinical pathology. Volume 150(2018)Supplement 1
• Page Start: S164
• Page End: S165
• Publication Date: 2018-09-21
• Subjects: Diagnosis, Laboratory -- Periodicals; Pathology -- Periodicals; 616.07
• DOI: 10.1093/ajcp/aqy112.381
• Languages: English
• ISSNs: 0002-9173

Abstract

Abstract: Introduction: Cytogenetic abnormalities, detectable by FISH and karyotype, can predict disease characteristics and inform prognosis in patients with plasma cell neoplasms. Our hospital sends bone marrow specimens with suspected or confirmed plasma cell neoplasms to a reference laboratory for plasma cell disorder FISH testing. FISH results are not reported on specimens with insufficient concentrations of plasma cells, as determined by screening with IGH probes. Approximately 45% of samples from our laboratory are insufficient for FISH scoring. The hospital is charged for these specimens nonetheless. We undertook a quality improvement initiative to identify pretest criteria that could predict whether a sample would be sufficient for FISH testing and that we could translate into an algorithm to determine which specimens to send for FISH. Methods: We analyzed all bone marrow specimens on which plasma cell FISH was requested from August 2015 through May 2017 (n = 437 samples, 220 of which were sufficient for FISH). Since all samples sent for FISH testing were evaluated by a hematopathologist at our hospital, we used corresponding pathology reports as the primary source of predictor data. We extracted features, including the percentages of plasma cells identified in the bone marrow core biopsy and aspirate count and percentages of clonal and polyclonal plasma cells enumerated by flow cytometry. Our dependent variable was dichotomous and indicated whether FISH results were reported to be sufficient or insufficient by the reference lab. We evaluated both the univariate discriminative power of each predictor variable using the area under the receiver operating characteristic curve (AUC) and a multivariate decision tree-based machine learning model. We calculated AUC values in R and trained and evaluated decision trees using the R "Rpart" package. Results and Discussion: The two most powerful predictors in our univariate analysis were clonal plasma cell percentage by flow cytometry and percentage of plasma cells on aspirate count with AUC values of 0.87 and 0.89, respectively. A multivariate decision tree predicted that specimens with ≥0.35% clonal plasma cells by flow cytometry and ≥6.5% total plasma cells by aspirate count would be adequate for FISH analysis. This decision rule had a sensitivity and specificity of 59% and 94%, respectively. While implementing this decision rule would significantly reduce the number of samples with insufficient FISH results, it would also lead us to hold back an unacceptable number of specimens that could provide meaningful FISH results despite low plasma cell numbers. Therefore, we continue to send all initial diagnostic marrow samples from patients with plasma cell neoplasms for FISH testing regardless of plasma cell percentage and are investigating alternative strategies to apply our findings and reduce unnecessary FISH testing.