Capi-score: a quantitative algorithm for identifying disease patterns in nailfold videocapillaroscopy

Abstract

Objectives: EULAR supports the use of nailfold videocapillaroscopy (NVC) for identifying disease patterns (DPs) associated with SSc and RP. Recently, EULAR proposed an easy-to-manage procedure, a so-called Fast Track algorithm, for differentiating SSc patterns from non-SSc patterns in NVC specimens. However, subjectivity among capillaroscopists remains a limitation. Our aim was to perform a software-based analysis of NVC peculiarities in a cohort of samples from SSc and RP patients and, subsequently, build a Fast Track-inspired algorithm for identifying DPs without the constraint of interobserver variability.

Methods: NVCs were examined by 9 capillaroscopists. Those NVCs whose DPs were consensually agreed upon (by ≥2 out of 3 interobservers) were subsequently analysed using in-house-developed software. The results for each variable were grouped according to the consensually agreed-upon DPs in order to identify useful hallmarks for categorizing them.

Results: A total of 851 NVCs (21 957 images) whose DPs had been consensually agreed upon were software-analysed. Appropriate cut-offs set for capillary density and percentage of abnormal and giant capillaries, tortuosities and haemorrhages allowed DP categorization and the development of the CAPI-score algorithm. This consisted of four rules: Rule 1, SSc vs non-SSc, accuracy 0.88; Rules 2 and 3, SSc-early vs SSc-active vs SSc-late, accuracy 0.82; Rule 4, non-SSc normal vs non-SSc non-specific, accuracy 0.73. Accuracy improved when the analysis was limited to NVCs whose DPs had achieved full consensus between the interobservers.

Conclusion: The CAPI-score algorithm may become a tool that is useful in assigning DPs by overcoming the limitations of subjectivity.

Keywords: Fast Track algorithm; Raynaud’s phenomenon; nailfold videocapillaroscopy; quantitative; software-based algorithm; systemic sclerosis.

Figure 1.

Flowchart diagram of the study. The sequence of steps in the study is shown, starting with the initially recruited NVCs and finishing with those finally selected to be software-analysed, which enabled us to develop a quantitative Fast Track–based algorithm. NS: non-specific; NVCs: nailfold videocapillaroscopies

Figure 2.

Confusion matrix showing matches and discrepancies between true and predicted disease patterns. All NVCs with consensus among capillaroscopists (gold standard: ≥2 out of 3 examiners, n = 851) were selected, and the proportion of matches (main diagonal of each matrix) and discrepancies (cells not on the main diagonal) when pattern assignment was performed by the gold standard (real patterns) or by software-based rules (predicted patterns) was annotated. Match percentage is indicated for each disease pattern. There were 30 capillaroscopies that were not analysed by the software, because they consisted of fewer than eight images or were not calibrated. NVCs, nailfold videocapillaroscopies

Figure 3.

CapiScore algorithm for discriminating between SSc disease patterns. After grouping by disease pattern, all NVCs with consensus among capillaroscopists (≥2 out of 3 examiners, n = 881) were examined using the software. Thirty of them were discarded because they consisted of fewer than eight images or were not calibrated. Taking the remaining 851 capillaroscopies, cut-offs of variables that best discriminated between patterns were selected, and a series of rules for categorization were proposed: first, between SSc and Non-SSc patterns; second, within SSc, between SSc-early, SSc-active and SSc-late patterns; finally, within Non-SSc, between Normal and Non-specific patterns. cap, capillaries; NVCs, nailfold videocapillaroscopies