Randomized controlled trial of a computer-assisted learning program to improve caries detection from bitewing radiographs

Abstract

Objectives: To investigate whether using a computer-assisted learning (CAL) calibration program improves the accuracy of dental students in caries detection from bitewing radiographs.

Methods: Dental students were assigned to an experimental (n=33) and control (n=34) group using a randomized block design. The experimental group used the CAL program with feedback to calibrate themselves against experts in radiographic caries detection and a histological gold standard of the actual clinical condition. Feedback was provided visually of the actual tooth surface condition and in the form of graphics showing diagnostic accuracy performance measures. Performance was tested before the program (for the control group) and after the program (for the experimental group) by assessing surfaces (n=56) from a new independent digital test set of evaluation radiographs for the presence, histologically, of dentine caries (n=23). The summary receiver operating characteristic (SROC) method for summarizing true positive ratio (TPR) (sensitivity) and false positive ratio (FPR) (1-specificity) was used to analyse the dichotomous data. Differences between the control and experimental groups were tested for (a) the area under the SROC curve (A(z)) and (b) the TPR, FPR and diagnostic odds ratio (DOR) using the Mann-Whitney test (P<0.05).

Results: The mean TPR for dentine caries detection was 76.3% (SD 13.0%) for the experimental group and 66.9% (SD 14.8%) for the control group (P=0.005). Mean FPRs were similar (experimental 28.1% and control 28.7%; P>0.5). The area under the SROC curve A(z) was 0.832 for the experimental group and 0.773 for the control group (P=0.002). The mean DOR for dentine caries in the experimental group (12.4) was better than that in the control group (8.8) (P=0.003).

Conclusions: The CAL program does improve diagnostic performance. Improving the cognitive feedback provided by the program should be considered before implementation.