Clinical Deployment of Interpretable AI: Bridging Routine Clinical Tests and Proteomic Signatures for Preeclampsia Risk Stratification.

INTRODUCTION

Preeclampsia (PE) is the second-leading global cause of maternal mortality, affecting 5% of primigravidas. Owing to the substantial heterogeneity of clinical manifestations in PE, an urgent need arises to quantitatively evaluate the efficacy of existing diagnostic methods based on positive proteinuria (PRO) and to develop novel biomarkers to enhance diagnostic accuracy.

METHODS

We based 1,215 pregnant women obtained from who delivery at the hospital in January 2018 and April 2022 and involved predictors of 66 routine clinical laboratory tests (RCLTs). In addition, from 362 peripheral blood proteomic samples obtained from published datasets. Compared, evaluated, and explored the performances of 5 machine learning models to constructed prediction models.

RESULTS

We pioneered the application of machine learning to assess the diagnostic efficiency of PRO quantitatively, AUROC of 0.771. Next, a more comprehensive assessment was discussed, including 66 RCTIs from blood and urine test items, the AUROC increased to 0.920. Furthermore, the feature selection strategy trained a superior routine clinical prediction model with 5 RCLTs (PRO, alkaline phosphatase (ALP), amylase (AMY), Uric Acid (UA), and Lactate Dehydrogenase (LDH)) for PE to ensure practicality and high performance. In addition, we constructed a protein prediction model for PE based on peripheral blood proteome. Subsequently, EphA1 has been identified as a protein candidate marker for PE, and is highly expressed in placentals. Finally, we established a user-friendly and interpretable PE risk prediction webserver (http://bioinfor. imu.edu.cn/lbppe/) to assist improve the PE diagnosis efficiency.

DISCUSSION: The predictive platform developed in this study enhances PE early detection, addressing the clinical need for rapid screening tools. Future multi-center trials should validate the models' generalizability.

CONCLUSION

This study assessed the diagnostic efficiency of proteinuria quantitatively and constructed a cost-effective PE prediction system, which is crucial for improving the diagnostic accuracy of PE.