The artificial intelligence (AI) system containing autoverification and autoguidance subsystems for urine microalbuminuria (Uma) detection is established in this study to not only enhance laboratory physician efficiency but also monitor the progression of diabetes kidney disease (DKD). The Uma detection includes Uma, urine creatinine and urine albumin-creatinine ratio (ACR) items. After verification of the detection results through the autoverification subsystem, a series of recommendations on diagnosis or further disposal of DKD to the clinician is put forward by the autoguidance subsystem.
A total of 351 clinical cases was used to establish the autoverification subsystem for Uma, which included the rules of the switch, instrument alarm, samples, quality control, detection limits, and consistency check according to the AUTO10-A document issued by the American Clinical Laboratory Standards Institute (1). Meanwhile, nine recommendations for the autoguidance subsystem were established according to the diagnostic guides of DKD (2,3) and diabetes (4). Recommendations will be provided with one or more of the following: R1) Possible DKD, please combine with fundus examination. R2) Consider other diseases leading to proteinuria after excluding diabetes history. R3) Consider other diseases leading to proteinuria. R4) Please check kidney function. R5) If the patient has a history of diabetes, please keep supervising ACR; otherwise, please combine with other relevant examinations. R6) Please recheck the routine urine test (RUT) and ACR a few days later. R7) Please perform the RUT to eliminate acute urinary tract inflammation. R8) Glucose was ≥7 mmol/L or/and hemoglobin A1c ≥6.5% at least once; if the patient has diabetes, please include fundus examination. And R9) Glucose was in the range from 6.10 to 6.99 mmol/L; if the patient has not been diagnosed with diabetes, please perform an oral glucose tolerance test to exclude diabetes. Please combine with fundus examination if the results meet the diagnostic criteria for diabetes. (The flowchart of the AI system is shown in Fig. 1.) After the system was established, 3,666 clinical cases in 2019 were used to verify the feasibility and accuracy of the system through clinical data validation as well as comparison between AI results and laboratory physician judgements. Then 305 autoguidance clinical cases were selected randomly for clinical follow-up until May 2021. The last diagnosis and clinical history were searched for the patients for R1, R2, R3, R5, R8, and R9, and the follow-up test reports were searched for the patients for R4, R6, and R7. The pass rate, accuracy, and consistency were valued for the autoverification subsystem, and the accuracy, consistency, and coincidence rate were used for the autoguidance subsystem. The consistency was calculated with the Cohen κ test.
The passing rate of the autoverification subsystem was 62.88% with high consistency (0.987, P < 0.001) and accuracy (99.40%), whereas the accuracy of the auto-guidance subsystem was 94.62% with high consistency (0.983, P < 0.001). The validation results showed that our autoverification subsystem could enhance laboratory physicians’ efficiency. Among the 305 follow-up patients, those with diabetes comprised 77.43%, and 88.11% of the patients with diabetes had DKD or had a medical history of diabetes over 10 years, indicating they could develop DKD (2,3). The coincidence rate of the whole autoguidance subsystem was 79.87% with the coincidence rates of R1, R2, R3, R5, and R8 being 82.00%, 68.42%, 87.50%, 88.00%, and 57.89%, respectively. In cases of R1, 42.00% had DKD and 40% had a medical history of diabetes for >10 years, indicating that R1 can be useful to draw a clinician’s attention to DKD. In advising the clinician to consider medical history, R2, R3, and R5 offer different disposal methods. R4 can suggest that clinicians test serum creatinine for those without results of serum creatinine for whom ACR ≥30 mg/g occurs only once. For R4, 88.00% had diabetes and 93.18% of those with diabetes had a normal glomerular filtration rate, representing the cohorts without DKD. R6 and R7 were established to remind the clinicians to exclude the acute phase changes of the urinary system. The results of case follow-up revealed that only 33.00% of patients were received RUT. Moreover, up to 69.70% of these patients showed indications of acute phase changes of the urinary system. Respectively, R8 and R9 are designed for patients whose results reach the criteria for diagnosis but are without diagnosis of diabetes and patients with the borderline findings.
In conclusion, the results presented in this small observational study provide evidence that the AI system for Uma detection can enhance work efficiency for laboratory physicians by autoverification and monitor the progression of DKD through offering different specific recommendations to the clinicians in the report.
X.L. and Y.X. contributed equally to this work.
Acknowledgments. The authors acknowledge the participation of the software engineer, physicians, and other staff in this study.
Duality of Interest. No potential conflicts of interest relevant to this article were reported.
Author Contributions. Q.Z. was responsible for study concept, design, and guidance. X.L. and Y.X. were responsible for flowchart and rules design of the AI system. X.L. and Y.T. were responsible for validation of the system. Y.T., X.Y., Y.W., W.X., W.G., and Q.W. were responsible for acquisition of data. Y.X., Y.T., B.C., and X.J. were responsible for statistical analysis. X.L., Y.X., and Q.Z., wrote and conducted critical revision of the manuscript. Q.Z. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.