Influences on accuracy of automated image-based estimation of neonatal serum bilirubin level using smartphone application under different circumstances

Objective To study the influence factors on the accuracy of a smartphone software application (APP) , which is an automated image-based bilirubin (AIB) testing technique for neonatal bilirubin. Methods This was a prospective study involving 179 jaundiced neonates admitted to the Department of Neonatology of the Affiliated Xuzhou Hospital of Southeast University from August to December, 2017. If blood sampling was required to determine total serum bilirubin (TSB), the transcutaneous bilirubin (TcB) and AIB on the sternum (glabella or the visible darkest yellow area were tested in 36 cases at the same time) were also tested. The concentrations of AIB were measured using a mobile phone-based neonatal jaundice monitoring APP with images taken by OPPO R11 smartphone (55 cases were also taken by Huawei Mate 8 and iPhone 6). The accuracy and non-inferiority of AIB comparing with TcB, the correlation and consistency between AIB and TSB and the value of AIB in predicting TSB were analyzed. Non-inferiority trial, student's t test, variance analysis, Pearson's correlation analysis, Bland-Altman plots and receiver operating characteristic (ROC) curve were used as statistical methods. Results A total of 179 neonates were enrolled in this study with the mean gestational age of (36.2±2.1) weeks and the mean birth weight of (2 871±735) g. Based on all data, the mean difference between the absolute value of AIB minus TSB (|AIB-TSB|) and the absolute value of TcB minus TSB (|TcB-TSB|) was 0.77 mg/dl (1 mg/dl=17.1 μmol/L) with the 95%CI of (0.60-0.95) mg/dl, suggesting that the accuracy of AIB was not inferior to that of the TcB. Furthermore, AIB was not inferior to TcB in its accuracy in different subgroups based upon the time after birth (≤48 h and >48 h), gestational age (≤35 weeks and >35 weeks) and in the daytime subgroup. In the nighttime subgroup, the mean difference between∣AIB-TSB∣and∣TcB-TSB∣was 1.47 mg/dl [95%CI: (1.08-1.87) mg/dl], which indicated that the accuracy of AIB was inferior to that of the TcB. There were good correlation (r=0.788) and highly consistency between AIB and TSB based on all data [96.4% (238/247) of the samples were within the 95% limits of agreement of (-4.75 to 5.71) mg/dl]. Such results were also found in different subgroups, such as smartphone brands, detection areas, the time after birth and gestational age. The correlation and consistency between AIB and TSB in daytime subgroup (r=0.924, 98.4%) were obviously stronger than those of the nighttime subgroup (r=0.727, 87.5%). The mean difference between the absolute value of ocular value minus TSB (|ocular value-TSB|) and |AIB-TSB| was 2.13 mg/dl [95%CI: (1.68-2.58) mg/dl], implying a superior accuracy of AIB than ocular estimation. The areas under the ROC curves, the sensitivity and the specificity of AIB for the prediction of TSB >10, >15 and >20 mg/dl were 0.94, 93% and 85%; 0.89, 75% and 87%; and 0.84, 50% and 88%, respectively. Conclusions AIB is not inferior to TcB in accuracy, and is significantly superior to ocular estimation. There are good correlation and strong consistency between AIB and TSB. The accuracy of AIB, and the correlation and consistency between AIB and TSB are less likely to be affected by the time after birth, gestational age, smartphone brands and areas being examined, but are largely affected when examining during nighttime. Therefore, it is recommended to detect AIB at daytime with bright natural light. Key words: Jaundice, Neonatal; Bilirubin; Reference values; Smartphone; Software; Image processing, computer-assisted