Artificial intelligence and amniotic fluid multiomics: prediction of perinatal outcome in asymptomatic women with short cervix.

OBJECTIVE: To evaluate the application of artificial intelligence (AI), i.e. deep learning and other machine-learning techniques, to amniotic fluid (AF) metabolomics and proteomics, alone and in combination with sonographic, clinical and demographic factors, in the prediction of perinatal outcome in asymptomatic pregnant women with short cervical length (CL). METHODS: AF samples, which had been obtained in the second trimester from asymptomatic women with short CL (< 15 mm) identified on transvaginal ultrasound, were analyzed. CL, funneling and the presence of AF 'sludge' were assessed in all cases close to the time of amniocentesis. A combination of liquid chromatography coupled with mass spectrometry and proton nuclear magnetic resonance spectroscopy-based metabolomics, as well as targeted proteomics analysis, including chemokines, cytokines and growth factors, was performed on the AF samples. To determine the robustness of the markers, we used six different machine-learning techniques, including deep learning, to predict preterm delivery < 34 weeks, latency period prior to delivery < 28 days after amniocentesis and requirement for admission to a neonatal intensive care unit (NICU). Omics biomarkers were evaluated alone and in combination with standard sonographic, clinical and demographic factors to predict outcome. Predictive accuracy was assessed using the area under the receiver-operating characteristics curve (AUC) with 95% CI, sensitivity and specificity. RESULTS: Of the 32 patients included in the study, complete omics, demographic and clinical data and outcome information were available for 26. Of these, 11 (42.3%) patients delivered ≥ 34 weeks, while 15 (57.7%) delivered < 34 weeks. There was no statistically significant difference in CL between these two groups (mean ± SD, 11.2 ± 4.4 mm vs 8.9 ± 5.3 mm, P = 0.31). Using combined omics, demographic and clinical data, deep learning displayed good to excellent performance, with an AUC (95% CI) of 0.890 (0.810-0.970) for delivery < 34 weeks' gestation, 0.890 (0.790-0.990) for delivery < 28 days post-amniocentesis and 0.792 (0.689-0.894) for NICU admission. These values were higher overall than for the other five machine-learning methods, although each individual machine-learning technique yielded statistically significant prediction of the different perinatal outcomes. CONCLUSIONS: This is the first study to report use of AI with AF proteomics and metabolomics and ultrasound assessment in pregnancy. Machine learning, particularly deep learning, achieved good to excellent prediction of perinatal outcome in asymptomatic pregnant women with short CL in the second trimester. Copyright © 2018 ISUOG. Published by John Wiley & Sons Ltd.

Metabolomic identification of diagnostic serum-based biomarkers for advanced stage melanoma.

INTRODUCTION: Melanoma is a highly aggressive malignancy and is currently one of the fastest growing cancers worldwide. While early stage (I and II) disease is highly curable with excellent prognosis, mortality rates rise dramatically after distant spread. We sought to identify differences in the metabolome of melanoma patients to further elucidate the pathophysiology of melanoma and identify potential biomarkers to aid in earlier detection of recurrence. METHODS: Using 1H NMR and DI-LC-MS/MS, we profiled serum samples from 26 patients with stage III (nodal metastasis) or stage IV (distant metastasis) melanoma and compared their biochemical profiles with 46 age- and gender-matched controls. RESULTS: We accurately quantified 181 metabolites in serum using a combination of 1H NMR and DI-LC-MS/MS. We observed significant separation between cases and controls in the PLS-DA scores plot (permutation test p-value = 0.002). Using the concentrations of PC-aa-C40:3, DL-carnitine, octanoyl-L-carnitine, ethanol, and methylmalonyl-L-carnitine we developed a diagnostic algorithm with an AUC (95% CI) = 0.822 (0.665-0.979) with sensitivity and specificity of 100 and 56%, respectively. Furthermore, we identified arginine, proline, tryptophan, glutamine, glutamate, glutathione and ornithine metabolism to be significantly perturbed due to disease (p < 0.05). CONCLUSION: Targeted metabolomic analysis demonstrated significant differences in metabolic profiles of advanced stage (III and IV) melanoma patients as compared to controls. These differences may represent a potential avenue for the development of multi-marker serum-based assays for earlier detection of recurrences, allow for newer, more effective targeted therapy when tumor burden is less, and further elucidate the pathophysiologic changes that occur in melanoma.