Predicting Placenta Accreta Spectrum Disorder Through Machine Learning Using Metabolomic and Lipidomic Profiling and Clinical Characteristics.

OBJECTIVE

To perform metabolomic and lipidomic profiling with plasma samples from patients with placenta accreta spectrum (PAS) to identify possible biomarkers for PAS and to predict PAS with machine learning methods that incorporated clinical characteristics with metabolomic and lipidomic profiles.

METHODS

This was a multicenter case-control study of patients with placenta previa with PAS (case group n=33) and previa alone (control group n=21). Maternal third-trimester plasma samples were collected and stored at -80°C. Untargeted metabolomic and targeted lipidomic assays were measured with flow-injection mass spectrometry. Univariate analysis provided an association of each lipid or metabolite with the outcome. The Benjamini-Hochberg procedure was used to control for the false discovery rate. Elastic net machine learning models were trained on patient characteristics to predict risk, and an integrated elastic net model of lipidome or metabolome with nine clinical features was trained. Performance using the area under the receiver operating characteristic curve (AUC) was determined with Monte Carlo cross-validation. Statistical significance was defined at P<.05.

RESULTS

The mean gestational age at sample collection was 33 3/7 weeks (case group) and 35 5/7 weeks (control group) (P<.01). In total, 786 lipid species and 2,605 metabolite features were evaluated. Univariate analysis revealed 31 lipids and 214 metabolites associated with the outcome (P<.05). After false discovery rate adjustment, these associations no longer remained statistically significant. When the machine learning model was applied, prediction of PAS with only clinical characteristics (AUC 0.685, 95% CI, 0.65-0.72) performed similarly to prediction with the lipidome model (AUC 0.699, 95% CI, 0.60-0.80) and the metabolome model (AUC 0.71, 95% CI, 0.66-0.76). However, integration of metabolome and lipidome with clinical features did not improve the model.

CONCLUSION

Metabolomic and lipidomic profiling performed similarly to, and not better than, clinical risk factors using machine learning to predict PAS among patients with PAS with previa and previa alone.