Blind Test Data Research Articles

Classification of Vaginal Cleanliness Grades through Surface‐Enhanced Raman Spectral Analysis via The Deep‐Learning Variational Autoencoder–Long Short‐Term Memory Model

In this study, it is aimed to establish a novel method based on a deep‐learning‐guided surface‐enhanced Raman spectroscopy (SERS) technique to achieve rapid and accurate classification of vaginal cleanliness levels. We proposed a variational autoencoder (VAE) approach to enhance spectral quality, coupled with a deep learning algorithm long short‐term memory (LSTM) neural network to analyze SERS spectra produced by vaginal secretions. The performance of various machine learning (ML) algorithms is assessed using multiple evaluation metrics. Finally, the reliability of the optimal model is tested using blind test data (N = 10/group for each cleanliness level). The data quality of the SERS fingerprints of four types of vaginal secretions is significantly improved after VAE decoding and reconstruction. The signal‐to‐noise ratio of the generated spectra increased from the original 2.58–11.13. Among all algorithms, the VAE–LSTM algorithm demonstrates the best prediction ability and time efficiency. Additionally, blind test datasets yielded an overall accuracy of 85%. In this study, it is concluded that the deep‐learning‐guided SERS technique holds significant potential in rapidly distinguishing between different levels of vaginal cleanliness through human vaginal secretion samples. This contributes to the efficient diagnosis of vaginal cleanliness levels in clinical settings.

GPU accelerated sparse curvelet-constrained wavefield reconstruction inversion with source estimation: Application to Chevron Benchmark 2014 blind test data set

Full-waveform inversion (FWI) is a pivotal tool in seismic exploration and seismology but encounters a persistent challenge known as “cycle-skipping,” causing it to frequently converge to local minima. Wavefield reconstruction inversion (WRI) has emerged as a potential solution to mitigate the risk of cycle skipping. Researchers have provided numerical examples indicating that WRI is less prone to being ensnared in local minima caused by the absence of low-frequency component data and the absence of a well-defined initial model compared with conventional FWI. Despite its potential, the computational demands of WRI have hindered its widespread application, especially in scenarios involving ocean towed streamer seismic acquisition and unknown sources, where the augmented systems differ from source to source. In our study, we introduce a novel approach — sparse curvelet-constrained WRI with source estimation (WRI-SE-CC) — accelerated by graphics processing unit (GPU). Real-time source function estimation is achieved through the variable projection method, and noise-related artifacts are suppressed using sparse curvelet constraints. By optimizing the utilization of hundreds of computation processors within a GPU for parallel computing of matrix-vector multiplications, we present a GPU-based grouped conjugate gradient method to accelerate the computation of WRI-SE-CC. Numerical experiments demonstrate a significant 240-fold acceleration compared with the preconditioned conjugate gradient using one CPU core for computations involving multiple sources. Inversion experiments with the overthrust model demonstrate the capability of our method in mitigating local minima and suppressing noise-related artifacts. Furthermore, we validate the framework on the Chevron 2014 blind test data set, showcasing its effectiveness in addressing practical challenges in the field.

Well-Log-Based Reservoir Property Estimation With Machine Learning: A Contest Summary

Well logs are processed and interpreted to estimate in-situ reservoir properties, which are essential for reservoir modeling, reserve estimation, and production forecasting. While the traditional methods are mostly based on multimineral physics or empirical formulae, machine learning provides an alternative data-driven approach that requires much less a-priori geological or petrophysical information. From October 2021 to March 2022, the Petrophysical Data-Driven Analytics Special Interest Group (PDDA SIG) of the Society of Petrophysicists and Well Log Analysts (SPWLA) hosted a machine-learning contest aiming to develop data-driven models for estimating reservoir properties, including shale volume, porosity, and fluid saturation, based on a common set of well logs, including gamma ray, bulk density, neutron porosity, resistivity, and sonic. Log data from nine wells from the same field, together with the interpreted reservoir properties by petrophysicists, were provided as training data, and five additional wells were provided as blind test data. During the contest, various data-driven models were developed by the contestants to predict the three reservoir properties with the provided training data set. The top five performing models from the contest, on average, beat the performance of the benchmarked Random Forest model by 45% in the root-mean-square error (RMSE) score. In the paper, we will review these top-performing solutions, including their preprocessing techniques, feature engineering, and machine-learning models, and summarize their advantages and conditions.

Convolutional neural network and level-set spectral element method for ultrasonic imaging of delamination cavities in an anisotropic composite structure

We present a computational approach that incorporates a convolutional neural network (CNN) for detecting internal delamination in a layered 2D plane-strain anisotropic composite structure of transient elastodynamic fields. The two-dimensional spectral element method (SEM) is utilized to simulate the propagation of elastic waves in an orthotropic solid sandwiched by isotropic solids and their interaction with the internal delamination cavity. This work generates training data consisting of input-layer features (i.e., measured wave signals) and output-layer features (i.e., element types, such as void or regular, of all elements in a domain).To accelerate training data generation, we utilize explicit time integration (e.g., the Runge–Kutta scheme) coupled with an SEM wave solver. Applying the level-set method additionally avoids having to perform an expensive re-meshing process for every possible configuration of the delamination cavities during the data-generation phase. The CNN is trained to classify each element as a non-void or void element from the measured wave signals. Clusters of identified void elements reconstruct targeted cavities. Once our neural network is trained using synthetic data, we analyze how effectively the CNN performs on synthetic measurement data. To this end, we use blind test data from a third-party simulator that explicitly models the traction-free boundary of cavities for anisotropic materials without the application of the level-set method.Our numerical examples show that our approach can effectively detect the internal cavities in an anisotropic structure made of aluminum and carbon fiber-reinforced epoxy using the measured elastic waves without any prior information about the cavities’ locations, shapes, and numbers. The presented method can be extended into a more realistic 3D setting and utilized for the nondestructive test of various anisotropic composite structures.

Pseudo-Computed Tomography Generation from Noisy Magnetic Resonance Imaging with Deep Learning Algorithm

Purpose: Magnetic Resonance Imaging (MRI) applications offer superior soft tissue contrast compared with Computed Tomography (CT) for accurate radiotherapy planning although MRI images suffer from poor image quality and lack electron density for radiation dose calculation. The present study aims to use the Deep Learning (DL) approach to 1) enhance the quality of MRI images and 2) generate synthetic CT images using MRI images for more accurate radiotherapy planning.
 Materials and Methods: In this paper, the pix2pix Generative Adversarial Network was utilized to synthesize CT images from noisy MRI images of 20 arbitrarily patients with brain disease. The standard statistical measurements investigated the accuracy comparison of the modeled Hounsfield Unit (HU) value from MRI images and referenced CT of each patient. The famous quality metrics that were used to compare synthetic CTs and referenced CTs were the Mean Absolute eError (MAE), the structural similarity index (SSIM), and the Peak Signal-to-Noise Ratio (PSNR).
 Results: The higher quality measurements between the synthetic pseudo-CT and the referenced CT images as PSNR and SSIM should correlate with the lower MAE value. For the overall brain among blind test data, the measured peak signal-to-noise ratio, mean absolute error, and structural similarity index values were about 16.5, 28.13, and 93.46, respectively.
 Conclusion: The proposed method provides an acceptable level of statistical measurements computed on the Pseudo-CT and referenced CT, and it could be concluded that the p-CT can be implemented in radiotherapy treatment planning with acceptable accuracy.

PON-Fold: Prediction of Substitutions Affecting Protein Folding Rate.

Most proteins fold into characteristic three-dimensional structures. The rate of folding and unfolding varies widely and can be affected by variations in proteins. We developed a novel machine-learning-based method for the prediction of the folding rate effects of amino acid substitutions in two-state folding proteins. We collected a data set of experimentally defined folding rates for variants and used them to train a gradient boosting algorithm starting with 1161 features. Two predictors were designed. The three-class classifier had, in blind tests, specificity and sensitivity ranging from 0.324 to 0.419 and from 0.256 to 0.451, respectively. The other tool was a regression predictor that showed a Pearson correlation coefficient of 0.525. The error measures, mean absolute error and mean squared error, were 0.581 and 0.603, respectively. One of the previously presented tools could be used for comparison with the blind test data set, our method called PON-Fold showed superior performance on all used measures. The applicability of the tool was tested by predicting all possible substitutions in a protein domain. Predictions for different conformations of proteins, open and closed forms of a protein kinase, and apo and holo forms of an enzyme indicated that the choice of the structure had a large impact on the outcome. PON-Fold is freely available.

The Best Scenario for Geostatistical Modeling of Porosity in the Sarvak Reservoir in an Iranian Oil Field, Using Electrofacies, Seismic Facies, and Seismic Attributes

Summary The lateral and vertical variations in porosity significantly impact the reservoir quality and the volumetric calculations in heterogeneous reservoirs. With a case study from Iran’s Zagros Basin Sarvak reservoir in the Dezful Embayment, this paper aims to demonstrate an efficient methodology for distributing porosity. Four facies models (based on electrofacies analysis data and seismic facies) with different geostatistical algorithms were used to examine the effect of different facies types on porosity propagation. Both deterministic and stochastic methods are adopted to check the impact of geostatistical algorithms on porosity modeling in the static model. A total of 40 scenarios were run and validated for porosity distribution through a blind test procedure to check the reliability of the models. The study’s findings revealed high correlation values in the blind test data for all porosity realizations linked to seismic facies, ranging from 0.778 to 0.876. In addition, co-kriging to acoustic impedance (AI), as a secondary variable, increases the correlation coefficient in all related cases. Unlike deterministic algorithms, using stochastic methods reduces the uncertainty and causes the porosity model to have an identical histogram compared with the original data. This study introduced a comprehensive workflow for porosity distribution in the studied carbonate Sarvak reservoir, considering the electrofacies, and seismic facies, and applying different geostatistical algorithms. As a result, based on this workflow, simultaneously linking the porosity distribution to seismic facies, co-kriging to AI, and applying the sequential Gaussian simulation (SGS) algorithm result in the best spatial modeling of porosity.

Empirical Correlations for Predicting Flow Rates Using Distributed Acoustic Sensor Measurements, Validated with Wellbore and Flow Loop Data Sets

Summary Distributed acoustic sensing (DAS) is an emerging surveillance technology that is becoming increasingly popular in the oil and gas industry for real-time flow monitoring. However, there are limited studies that rigorously quantify flow rates using DAS. This work expands the existing literature by presenting a detailed workflow for accurately estimating fluid flow rates from DAS data using time- and frequency-domain signal processing. Three simple empirical correlation functions (linear, exponential, and cubic) are developed and tested to predict flow rates from DAS. The proposed correlations are demonstrated for flow rates ranging from 50 to 300 gallons per minute (GPM) in a vertical 5,163-ft-deep wellbore and from 12 to 36 GPM in a horizontal surface flow loop. Tests were performed using a single-phase flow of water as well as using synthetic oil-based drilling mud. Time-domain DAS processing using root-mean-square (RMS) value and frequency-domain processing using frequency band energy (FBE) is evaluated, followed by a statistical approach to minimize the influence of outliers. The RMS and FBE approaches are individually compared for flow prediction, and the performance of the correlations is rigorously evaluated on a blind data set that was not originally used for developing the correlations. For both the wellbore and flow loop data sets, a coefficient of determination (or R2) greater than 0.95 with an average flow rate prediction error of less than 10% was achieved for the best-performing correlation for the blind test data. The analysis procedure and workflow presented in this study can be adopted and extended to different operating conditions for quantitative flow rate prediction using DAS.

Level-Set and Learn: Convolutional Neural Network for Classification of Elements to Identify an Arbitrary Number of Voids in a 2D Solid Using Elastic Waves

We present a new convolutional neural network (CNN)-based element-wise classification method to detect a random number of voids with arbitrary shapes in a two-dimensional (2D) plane-strain solid subjected to elastodynamics. We consider that an elastic wave source excites the solid including a random number of voids, and wave responses are measured by sensors placed around the solid. We present a CNN for resolving the inverse problem, which is formulated as an element-wise classification problem. The CNN is trained to classify each element into a regular or void element from measured wave signals. Element-wise binary classification enables the identification of targeted voids of any shapes and any number without prior knowledge or hint about their locations, shape types, and numbers, while existing methods rely on such prior information. To this end, we generate training data consisting of input-layer features (i.e., measured wave signals at sensors) and output-layer features (i.e., element types of all elements). When the training data are generated, we utilize the level-set method to avoid an expensive remeshing process, which is otherwise needed for each different configuration of voids. We also analyze how effectively the CNN performs on blind test data from a non-level-set wave solver that explicitly models the boundary of voids using an unstructured, fine mesh. Numerical results show that the suggested approach can detect the locations, shapes, and sizes of multiple elliptical and circular voids in the 2D solid domain in the test data set as well as a blind test data set.

Korea Pathfinder Lunar Orbiter Flight Dynamics Simulation and Rehearsal Results for Its Operational Readiness Checkout

Korea Pathfinder Lunar Orbiter (KPLO), also known as Danuri, was successfully launched on 4 Aug. from Cape Canaveral Space Force Station using a Space-X Falcon-9 rocket. Flight dynamics (FD) operational readiness was one of the critical parts to be checked before the flight. To demonstrate FD software’s readiness and enhance the operator’s contingency response capabilities, KPLO FD specialists planned, organized, and conducted four simulations and two rehearsals before the KPLO launch. For the efficiency and integrity of FD simulation and rehearsal, different sets of blind test data were prepared, including the simulated tracking measurements that incorporated dynamical model errors, maneuver execution errors, and other errors associated with a tracking system. This paper presents the simulation and rehearsal results with lessons learned for the KPLO FD operational readiness checkout. As a result, every functionality of FD operation systems is firmly secured based on the operation procedure with an enhancement of contingency operational response capability. After conducting several simulations and rehearsals, KPLO FD specialists were much more confident in the flight teams’ ability to overcome the challenges in a realistic flight and FD software’s reliability in flying the KPLO. Moreover, the results of this work will provide numerous insights to the FD experts willing to prepare deep space flight operations.

ProTstab2 for Prediction of Protein Thermal Stabilities.

The stability of proteins is an essential property that has several biological implications. Knowledge about protein stability is important in many ways, ranging from protein purification and structure determination to stability in cells and biotechnological applications. Experimental determination of thermal stabilities has been tedious and available data have been limited. The introduction of limited proteolysis and mass spectrometry approaches has facilitated more extensive cellular protein stability data production. We collected melting temperature information for 34,913 proteins and developed a machine learning predictor, ProTstab2, by utilizing a gradient boosting algorithm after testing seven algorithms. The method performance was assessed on a blind test data set and showed a Pearson correlation coefficient of 0.753 and root mean square error of 7.005. Comparison to previous methods indicated that ProTstab2 had superior performance. The method is fast, so it was applied to predict and compare the stabilities of all proteins in human, mouse, and zebrafish proteomes for which experimental data were not determined. The tool is freely available.

The edge-guided FPN model for automatic stratigraphic correlation of well logs

The stratigraphic correlation of well logs is a fundamental and important topic in seismic well log interpretation. As the number of well logs increases, the automatic stratigraphic correlation of well logs is difficult and time-consuming. With the development of the deep learning (DL) models, several of them have been applied to solve the stratigraphic correlation. However, most of these DL models do not consider the multi-scale features of well logs, which makes against with the accurate and automatic stratigraphic correlation. In this study, by utilizing the multi-scale Feature Pyramid Network (FPN) model, we propose an edge-guided FPN (EFPN) model for the automatic stratigraphic correlation. The proposed EFPN model first extracts the multi-scale features of well logs, which benefits for the automatic stratigraphic correlation. Then, the EFPN model integrates the multi-scale features via the concat operation, where a side output is added for the further edge guidance. Moreover, we propose a hybrid loss function by combining the edge loss and the cross entropy loss for the automatic stratigraphic correlation. Based on the well logs from Daqing Oilfield, China, we first divide the well logs into the training data set and the blind test data set. After training the EFPN model, we make detailed comparisons with the widely used SegNet model. The numerical results show that the proposed EFPN model could obtain the stratigraphic correlation of well logs more accurately than the SegNet model, especially for the stratigraphic sequence boundaries and the boundaries of well logs.

Prediction of Casing Collapse Strength Based on Bayesian Neural Network

With the application of complex fracturing and other complex technologies, external extrusion has become the main cause of casing damage, which makes non-API high-extrusion-resistant casing continuously used in unconventional oil and gas resources exploitation. Due to the strong sensitivity of string ovality, uneven wall thickness, residual stress, and other factors to high anti-collapse casing, the API formula has a big error in predicting the anti-collapse strength of high anti-collapse casing. Therefore, Bayesian regularization artificial neural network (BRANN) is used to predict the external collapse strength of high anti-collapse casing. By collecting full-scale physical data, including initial defect data, geometric size, mechanical parameters, etc., after data preprocessing, the casing collapse strength data set is established for model training and blind measurement. Under the classical three-layer neural network, the Bayesian regularization algorithm is used for training. Through empirical formula and trial and error method, it is determined that when the number of hidden neurons is 12, the model is the best prediction model for high collapse resistance casing. The prediction results of the blind test data imported by the model show that the coincidence rate of BRANN casing collapse strength prediction can reach 96.67%. Through error analysis with API formula prediction results and KT formula prediction results improved by least square fitting, the BRANN-based casing collapse strength prediction has higher accuracy and stability. Compared with the traditional prediction method, this model can be used to predict casing strength under more complicated working conditions, and it has a certain guiding significance.

Real vs. simulated: Questions on the capability of simulated datasets on building fault detection for energy efficiency from a data-driven perspective

Literature on building Automatic Fault Detection and Diagnosis (AFDD) mainly focuses on simulated system data due to high expenses and difficulties of obtaining and analyzing real building data. There is a lack of validation on performances and scalabilities of data-driven AFDD approaches using simulated data and how it compares to that from real building data. In this study, we conduct two sets of experiments to seek answers to this question. We first evaluate data-driven fault detection strategies on real and simulated building data separately. We observe that the fault detection performances are not affected by fault detection strategies, sizes of training data, and the number of cross-validation folds when training and blind test data come from the same data source, namely, simulated or real building data. Next, we conduct a cross-dataset study, that is, develop the model using simulated data and tested on real building data. The results indicate the model trained on simulated data is not generalized to be applied for real building data for fault detection. Kolmogorov-Smirnov Test is conducted to confirm that there exist statistical differences between the simulated and real building data and identify a subset of features with similarities between the two datasets. Using the subset of the feature, cross-dataset experiments show fault detection improvements on most fault cases. We conclude that even if the system produces simulated data with the same fault symptoms from physical analysis perspectives, not all features from simulated datasets may not be beneficial for AFDD but only a subset of features contains valuable information from a machine learning perspective.

Detection of self-harm and suicidal ideation in emergency department triage notes.

Accurate identification of self-harm presentations to Emergency Departments (ED) can lead to more timely mental health support, aid in understanding the burden of suicidal intent in a population, and support impact evaluation of public health initiatives related to suicide prevention. Given lack of manual self-harm reporting in ED, we aim to develop an automated system for the detection of self-harm presentations directly from ED triage notes. We frame this as supervised classification using natural language processing (NLP), utilizing a large data set of 477 627 free-text triage notes from ED presentations in 2012-2018 to The Royal Melbourne Hospital, Australia. The data were highly imbalanced, with only 1.4% of triage notes relating to self-harm. We explored various preprocessing techniques, including spelling correction, negation detection, bigram replacement, and clinical concept recognition, and several machine learning methods. Our results show that machine learning methods dramatically outperform keyword-based methods. We achieved the best results with a calibrated Gradient Boosting model, showing 90% Precision and 90% Recall (PR-AUC 0.87) on blind test data. Prospective validation of the model achieves similar results (88% Precision; 89% Recall). ED notes are noisy texts, and simple token-based models work best. Negation detection and concept recognition did not change the results while bigram replacement significantly impaired model performance. This first NLP-based classifier for self-harm in ED notes has practical value for identifying patients who would benefit from mental health follow-up in ED, and for supporting surveillance of self-harm and suicide prevention efforts in the population.

Towards sound based testing of COVID-19—Summary of the first Diagnostics of COVID-19 using Acoustics (DiCOVA) Challenge

The technology development for point-of-care tests (POCTs) targeting respiratory diseases has witnessed a growing demand in the recent past. Investigating the presence of acoustic biomarkers in modalities such as cough, breathing and speech sounds, and using them for building POCTs can offer fast, contactless and inexpensive testing. In view of this, over the past year, we launched the “Coswara” project to collect cough, breathing and speech sound recordings via worldwide crowdsourcing. With this data, a call for development of diagnostic tools was announced in the Interspeech 2021 as a special session titled “Diagnostics of COVID-19 using Acoustics (DiCOVA) Challenge”. The goal was to bring together researchers and practitioners interested in developing acoustics-based COVID-19 POCTs by enabling them to work on the same set of development and test datasets. As part of the challenge, datasets with breathing, cough, and speech sound samples from COVID-19 and non-COVID-19 individuals were released to the participants. The challenge consisted of two tracks. The Track-1 focused only on cough sounds, and participants competed in a leaderboard setting. In Track-2, breathing and speech samples were provided for the participants, without a competitive leaderboard. The challenge attracted 85 plus registrations with 29 final submissions for Track-1. This paper describes the challenge (datasets, tasks, baseline system), and presents a focused summary of the various systems submitted by the participating teams. An analysis of the results from the top four teams showed that a fusion of the scores from these teams yields an area-under-the-receiver operating curve (AUC-ROC) of 95.1% on the blind test data. By summarizing the lessons learned, we foresee the challenge overview in this paper to help accelerate technological development of acoustic-based POCTs.

Strategies to Enhance the Performance of Gaussian Mixture Model Fitting for Uncertainty Quantification

Summary When formulating history matching within the Bayesian framework, we may quantify the uncertainty of model parameters and production forecasts using conditional realizations sampled from the posterior probability density function (PDF). It is quite challenging to sample such a posterior PDF. Some methods [e.g., Markov chain Monte Carlo (MCMC)] are very expensive, whereas other methods are cheaper but may generate biased samples. In this paper, we propose an unconstrained Gaussian mixture model (GMM) fitting method to approximate the posterior PDF and investigate new strategies to further enhance its performance. To reduce the central processing unit (CPU) time of handling bound constraints, we reformulate the GMM fitting formulation such that an unconstrained optimization algorithm can be applied to find the optimal solution of unknown GMM parameters. To obtain a sufficiently accurate GMM approximation with the lowest number of Gaussian components, we generate random initial guesses, remove components with very small or very large mixture weights after each GMM fitting iteration, and prevent their reappearance using a dedicated filter. To prevent overfitting, we add a new Gaussian component only if the quality of the GMM approximation on a (large) set of blind-test data sufficiently improves. The unconstrained GMM fitting method with the new strategies proposed in this paper is validated using nonlinear toy problems and then applied to a synthetic history-matching example. It can construct a GMM approximation of the posterior PDF that is comparable to the MCMC method, and it is significantly more efficient than the constrained GMM fitting formulation (e.g., reducing the CPU time by a factor of 800 to 7,300 for problems we tested), which makes it quite attractive for large-scale history-matching problems. NOTE: This paper is also published as part of the 2021 SPE Reservoir Simulation Conference Special Issue.

Abstract 14086: A Hybrid Machine Learning Method for Instantaneous Classification of Left Ventricular Filling Pressure Using Femoral Waveforms

Introduction: Non-invasive and instantaneous evaluation of left ventricular end-diastolic pressure (LVEDP) is of high significance in diagnosis and treatment of heart failure patients. We recently proposed the cardiac triangle mapping (CTM), a systems approach that uses information from arterial pressure waveforms and pre-ejection periods to map the global ventricular function (Pahlevan et al. Fluids 4.1 (2019): 16). Here, we propose a hybrid machine learning (ML) method based on CTM theory that uses pressure waveforms at the iliac bifurcation level and ECG to instantaneously classify the level of LVEDP. Methods: We studied 46 patients (Age: 39-90 (66.4±9.9), BMI: 20.2-36.8 (27.6±4.1), 12 females) who were scheduled for clinical left heart catheterization or coronary angiogram at the Keck Medical Center of USC. Exclusion criteria were valvular heart disease, left bundle branch block, or atrial fibrillation. Invasive LVEDP and pressure waveforms at iliac bifurcation were measured using a 3F Millar transducer tipped catheter, with simultaneous 3 channel ECG. The range of patients’ LVEDP was from 9.3 to 40.5 mmHg. LVEDP=18 mmHg was used as the ML classification cut-off. A random forest classifier (of n=30 grown decision trees, ensemble learning method: bootstrap aggregation, 6 inputs) was trained using data from 36 patients. The model was blindly tested on 10 additional patients. Results: Our ML model showed accuracies of 100.0% and 80.0% for predicting the true class of LVEDP on blind test data (i.e., elevated or normal, respectively) (Fig1). The accuracy on all data was higher than 93.4% for all classes. Conclusions: We demonstrated the proof-of-concept that a physics-based ML model can instantaneously classify the LV filling pressure using information extracted from femoral waveform and ECG. Our study is invasive validation; however, all required ML input parameters can be obtained noninvasively and remotely (i.e., using wearable devices or a smartphone).

O-121 Exploring non-invasive methods to predict Ploidy Status: Combination of blastocyst morphology image analysis and proteomic profiles by using Artificial Neural Networks

Abstract Study question Is the blastocyst morphology image analysis combined with the protein content of spent embryo culture medium a suitable way to predict embryo ploidy? Summary answer Morphological variables from blastocyst image analysis combined with IL-6 or MMP-1 concentration in spent culture medium showed more than 80% of accuracy for euploidy prediction. What is known already An artificial intelligence model based on the proteomic profile of euploid embryos and morphological data from blastocyst time-lapse images has been recently published (Bori et al., 2020). The most promising artificial neural network (ANN) algorithm considered 20 morphological variables extracted from image analysis and two proteins detected in embryo culture medium (MMP-1 and IL-6). The overall success rate on blind test data was 72.7% for live birth prediction. The main aim of the present study was to check if the same morphological variables combined with MMP-1 or IL-6 with a cost-effective technique could discriminate between euploid and aneuploid embryos. Study design, size, duration This prospective study included 120 embryos from the preimplantation genetic testing for aneuploidies (PGT-A) program. A single blastocyst image was obtained for each embryo and their spent culture medium was collected on the day 5/6 of embryo development (day of trophectoderm biopsy). Morphological variables were extracted for all the blastocyst. On the other hand, we quantified IL-6 levels of 67 embryos and MMP-1 levels of 53 embryos. Resulting parameters were used to predict PGT-A results. Participants/materials, setting, methods Blastocyst images were imported into Matlab software and segmented into regions of interest. We obtained 20 mathematical variables related to measurements of areas, number of pixels and texture analysis. Chromosome analysis was performed using next-generation sequence technology. In parallel, 20 µL of spent culture medium from each blastocyst was analyzed with ELISA kits (IL-6 or MMP-1). Protein concentrations and morphological variables were used as input data for an ANN associated with genetic algorithms. Main results and the role of chance The euploid rate for the set of embryos included in the IL-6 group was 51.4%. The ANN was trained with 49 embryos and blind tested with 18 embryos. Following results correspond to euploidy prediction on the blind test. The sensitivity, specificity, accuracy and area under the ROC curve (AUC) were: 0.56, 0.78, 0.67 and 0.72 considering only IL-6 values; 0.88, 0.78, 0.83 and 0.61 considering IL-6 values and blastocyst morphological data extracted from the image analysis. The euploid rate for the set of embryos included in the MMP-1 group was 51.9%. The ANN was trained with 39 embryos and blind tested with 14 embryos. Following results correspond to euploidy prediction on the blind test. The sensitivity, specificity, accuracy and AUC were: 0.71, 0.57, 0.64 and 0.67 considering only MMP-1 values; 0.86, 0.86, 0.86 and 0.61 considering MMP-1 values and morphological data extracted from the image analysis. Limitations, reasons for caution The detection limit in protein quantification is the main limitation of our study. The small number of embryos and the specific culture medium used should be considered for the model application. Wider implications of the findings Our preliminary results showed that blastocyst morphology and embryo secretomics could be useful for euploidy prediction by using artificial intelligence techniques. These findings may contribute to the emerging era of non-invasive preimplantation genetic testing (ni-PGT-A). Trial registration number not applicable

P–165 Using Artificial Intelligence to Classify Embryo Shape: An International Perspective

Abstract Study question Is a pre-trained machine learning algorithm able to accurately detect cellular arrangement in 4-cell embryos from a different continent? Summary answer Artificial Intelligence (AI) analysis of 4-cell embryo classification is transferable across clinics globally with 79% accuracy. What is known already Previous studies observing four-cell human embryo configurations have demonstrated that non-tetrahedral embryos (embryos in which cells make contact with fewer than 3 other cells) are associated with compromised blastulation and implantation potential. Previous research by this study group has indicated the efficacy of AI models in classification of tetrahedral and non-tetrahedral embryos with 87% accuracy, with a database comprising 2 clinics both from the same country (Brazil). This study aims to evaluate the transferability and robustness of this model on blind test data from a different country (France). Study design, size, duration The study was a retrospective cohort analysis in which 909 4-cell embryo images (“tetrahedral”, n = 749; “non-tetrahedral”, n = 160) were collected from 3 clinics (2 Brazilian, 1 French). All embryos were captured at the central focal plane using Embryoscope™ time-lapse incubators. The training data consisted solely of embryo images captured in Brazil (586 tetrahedral; 87 non-tetrahedral) and the test data consisted exclusively of embryo images captured in France (163 tetrahedral; 72 non-tetrahedral). Participants/materials, setting, methods The embryo images were labelled as either “tetrahedral” or “non-tetrahedral” at their respective clinics. Annotations were then validated by three operators. A ResNet–50 neural network model pretrained on ImageNet was fine-tuned on the training dataset to predict the correct annotation for each image. We used the cross entropy loss function and the RMSprop optimiser (lr = 1e–5). Simple data augmentations (flips and rotations) were used during the training process to help counteract class imbalances. Main results and the role of chance Our model was capable of classifying embryos in the blind French test set with 79% accuracy when trained with the Brazilian data. The model had sensitivity of 91% and 51% for tetrahedral and non-tetrahedral embryos respectively; precision was 81% and 73%; F1 score was 86% and 60%; and AUC was 0.61 and 0.64. This represents a 10% decrease in accuracy compared to when the model both trained and tested on different data from the same clinics. Limitations, reasons for caution Although strict inclusion and exclusion criteria were used, inter-operator variability may affect the pre-processing stage of the algorithm. Moreover, as only one focal plane was used, ambiguous cases were interpoloated and further annotated. Analysing embryos at multiple focal planes may prove crucial in improving the accuracy of the model. Wider implications of the findings: Though the use of machine learning models in the analysis of embryo imagery has grown in recent years, there has been concern over their robustness and transferability. While previous results have demonstrated the utility of locally-trained models, our results highlight the potential for models to be implemented across different clinics. Trial registration number Not applicable