Precision-recall Curve Research Articles

A simple machine learning model for the prediction of acute kidney injury following noncardiac surgery in geriatric patients: a prospective cohort study

BackgroundSurgery in geriatric patients often poses risk of major postoperative complications. Acute kidney injury (AKI) is a common complication following noncardiac surgery and is associated with increased mortality. Early identification of geriatric patients at high risk of AKI could facilitate preventive measures and improve patient prognosis. This study used machine learning methods to identify important features and predict AKI following noncardiac surgery in geriatric patients.MethodsThe data for this study were obtained from a prospective cohort. Patients aged ≥ 65 years who received noncardiac surgery from June 2019 to December 2021 were enrolled. Data were split into training set (from June 2019 to March 2021) and internal validation set (from April 2021 to December 2021) by time. The least absolute shrinkage and selection operator (LASSO) regularization algorithm and the random forest recursive feature elimination algorithm (RF-RFE) were used to screen important predictors. Models were trained through extreme gradient boosting (XGBoost), random forest, and LASSO. The SHapley Additive exPlanations (SHAP) package was used to interpret the machine learning model.ResultsThe training set included 6753 geriatric patients. Of these, 250 (3.70%) patients developed AKI. The XGBoost model with RF-RFE selected features outperformed other models with an area under the precision-recall curve (AUPRC) of 0.505 (95% confidence interval [CI]: 0.369–0.626) and an area under the receiver operating characteristic curve (AUROC) of 0.806 (95%CI: 0.733–0.875). The model incorporated ten predictors, including operation site and hypertension. The internal validation set included 3808 geriatric patients, and 96 (2.52%) patients developed AKI. The model maintained good predictive performance with an AUPRC of 0.431 (95%CI: 0.331–0.524) and an AUROC of 0.845 (95%CI: 0.796–0.888) in the internal validation.ConclusionsThis study developed a simple machine learning model and a web calculator for predicting AKI following noncardiac surgery in geriatric patients. This model may be a valuable tool for guiding preventive measures and improving patient prognosis.Trial registrationThe protocol of this study was approved by the Committee of Ethics from West China Hospital of Sichuan University (2019–473) with a waiver of informed consent and registered at www.chictr.org.cn (ChiCTR1900025160, 15/08/2019).

Radiomic-based prediction of lesion-specific systemic treatment response in metastatic disease

Despite sharing the same histologic classification, individual tumors in multi metastatic patients may present with different characteristics and varying sensitivities to anticancer therapies. In this study, we investigate the utility of radiomic biomarkers for prediction of lesion-specific treatment resistance in multi metastatic leiomyosarcoma patients. Using a dataset of n=202 lung metastases (LM) from n=80 patients with 1648 pre-treatment computed tomography (CT) radiomics features and LM progression determined from follow-up CT, we developed a radiomic model to predict the progression of each lesion. Repeat experiments assessed the relative predictive performance across LM volume groups. Lesion-specific radiomic models indicate up to a 4.5-fold increase in predictive capacity compared with a no-skill classifier, with an area under the precision-recall curve of 0.70 for the most precise model (FDR = 0.05). Precision varied by administered drug and LM volume. The effect of LM volume was controlled by removing radiomic features at a volume-correlation coefficient threshold of 0.20. Predicting lesion-specific responses using radiomic features represents a novel strategy by which to assess treatment response that acknowledges biological diversity within metastatic subclones, which could facilitate management strategies involving selective ablation of resistant clones in the setting of systemic therapy.

Exploring potential neuroimaging biomarkers for the response to non-steroidal anti-inflammatory drugs in episodic migraine

BackgroundNon-steroidal anti-inflammatory drugs (NSAIDs) are considered first-line medications for acute migraine attacks. However, the response exhibits considerable variability among individuals. Thus, this study aimed to explore a machine learning model based on the percentage of amplitude oscillations (PerAF) and gray matter volume (GMV) to predict the response to NSAIDs in migraine treatment.MethodsPropensity score matching was adopted to match patients having migraine with response and nonresponse to NSAIDs, ensuring consistency in clinical characteristics and migraine-related features. Multimodal magnetic resonance imaging was employed to extract PerAF and GMV, followed by feature selection using the least absolute shrinkage and selection operator regression and recursive feature elimination algorithms. Multiple predictive models were constructed and the final model with the smallest predictive residuals was chosen. The model performance was evaluated using the area under the receiver operating characteristic (ROCAUC) curve, area under the precision-recall curve (PRAUC), balance accuracy (BACC), sensitivity, F1 score, positive predictive value (PPV), and negative predictive value (NPV). External validation was performed using a public database. Then, correlation analysis was performed between the neuroimaging predictors and clinical features in migraine.ResultsOne hundred eighteen patients with migraine (59 responders and 59 non-responders) were enrolled. Six features (PerAF of left insula and left transverse temporal gyrus; and GMV of right superior frontal gyrus, left postcentral gyrus, right postcentral gyrus, and left precuneus) were observed. The random forest model with the lowest predictive residuals was selected and model metrics (ROCAUC, PRAUC, BACC, sensitivity, F1 score, PPV, and NPV) in the training and testing groups were 0.982, 0.983, 0.927, 0.976, 0.930, 0.889, and 0.973; and 0.711, 0.648, 0.639, 0.667,0.649, 0.632, and 0.647, respectively. The model metrics of external validation were 0.631, 0.651, 0.611, 0.808, 0.656, 0.553, and 0.706. Additionally, a significant positive correlation was found between the GMV of the left precuneus and attack time in non-responders.ConclusionsOur findings suggest the potential of multimodal neuroimaging features in predicting the efficacy of NSAIDs in migraine treatment and provide novel insights into the neural mechanisms underlying migraine and its optimized treatment strategy.

Discrimination of multiple sclerosis using scanning laser ophthalmoscopy images with autoencoder-based feature extraction

ObjectiveOptical coherence tomography (OCT) investigations have revealed that the thickness of inner retinal layers becomes decreased in multiple sclerosis (MS) patients, compared to healthy control (HC) individuals. To date, a number of studies have applied machine learning to OCT thickness measurements, aiming to enable accurate and automated diagnosis of the disease. However, there have much less emphasis on other less common retinal imaging modalities, like infrared scanning laser ophthalmoscopy (IR-SLO), for classifying MS. IR-SLO uses laser light to capture high-resolution fundus images, often performed in conjunction with OCT to lock B-scans at a fixed position. MethodsWe incorporated two independent datasets of IR-SLO images from the Isfahan and Johns Hopkins centers, consisting of 164 MS and 150 HC images. A subject-wise data splitting approach was employed to ensure that there was no leakage between training and test datasets. Several state-of-the-art convolutional neural networks (CNNs), including VGG-16, VGG-19, ResNet-50, and InceptionV3, and a CNN with a custom architecture were employed. In the next step, we designed a convolutional autoencoder (CAE) to extract semantic features subsequently given as inputs to four conventional ML classifiers, including support vector machine (SVM), k-nearest neighbor (K-NN), random forest (RF), and multi-layer perceptron (MLP). ResultsThe custom CNN (85 % accuracy, 85 % sensitivity, 87 % specificity, 93 % area under the receiver operating characteristics [AUROC], and 94 % area under the precision-recall curve [AUPRC]) outperformed state-of-the-art models (84 % accuracy, 83 % sensitivity, 87 % specificity, 92 % AUROC, and 94 % AUPRC); however, utilizing a combination of the CAE and MLP yields even superior results (88 % accuracy, 86 % sensitivity, 91 % specificity, 94 % AUROC, and 95 % AUPRC). ConclusionsWe utilized IR-SLO images to differentiate between MS and HC eyes, with promising results achieved using a combination of CAE and MLP. Future multi-center studies involving more heterogenous data are necessary to assess the feasibility of integrating IR-SLO images into routine clinical practice.

DMLS: an automated pipeline to extract the Drosophila modular transcription regulators and targets from massive literature articles.

Transcription regulation in multicellular species is mediated by modular transcription factor (TF) binding site combinations termed cis-regulatory modules (CRMs). Such CRM-mediated transcription regulation determines the gene expression patterns during development. Biologists frequently investigate CRM transcription regulation on gene expressions. However, the knowledge of the target genes and regulatory TFs participating in the CRMs under study is mostly fragmentary throughout the literature. Researchers need to afford tremendous human resources to fully surf through the articles deposited in biomedical literature databases in order to obtain the information. Although several novel text-mining systems are now available for literature triaging, these tools do not specifically focus on CRM-related literature prescreening, failing to correctly extract the information of the CRM target genes and regulatory TFs from the literature. For this reason, we constructed a supportive auto-literature prescreener called Drosophila Modular transcription-regulation Literature Screener (DMLS) that achieves the following: (i) prescreens articles describing experiments on modular transcription regulation, (ii) identifies the described target genes and TFs of the CRMs under study for each modular transcription-regulation-describing article and (iii) features an automated and extendable pipeline to perform the task. We demonstrated that the final performance of DMLS in extracting the described target gene and regulatory TF lists of CRMs under study for given articles achieved test macro area under the ROC curve (auROC) = 89.7% and area under the precision-recall curve (auPRC) = 77.6%, outperforming the intuitive gene name-occurrence-counting method by at least 19.9% in auROC and 30.5% in auPRC. The web service and the command line versions of DMLS are available at https://cobis.bme.ncku.edu.tw/DMLS/ and https://github.com/cobisLab/DMLS/, respectively. Database Tool URL: https://cobis.bme.ncku.edu.tw/DMLS/.

The use of artificial intelligence to detect voided medication orders in oral and maxillofacial surgery inpatients.

The aim of this study is to determine if supervised machine learning algorithms can accurately predict voided computerized physician order entry in oral and maxillofacial surgery inpatients. Data from Electronic Medical Record included patient demographics, comorbidities, procedures, vital signs, laboratory values, and medication orders were retrospectively collected. Predictor variables included patient demographics, comorbidities, procedures, vital signs, and laboratory values. Outcome of interest is if a medication order was voided or not. Data was cleaned and processed using Microsoft Excel and Python v3.12. Gradient Boosted Decision Trees, Random Forest, K-Nearest Neighbor, and Naïve Bayes were trained, validated, and tested for accuracy of the prediction of voided medication orders. 37,493 medication orders from 1,204 patient admissions over 5 years were used for this study. 3,892 (10.4%) medication orders were voided. Gradient Boosted Decision Trees, Random Forest, K-Nearest Neighbor, and Naïve Bayes had an Area Under the Receiver Operating Curve of 0.802with 95% CI [0.787, 0.825], 0.746 with 95% CI [0.722, 0.765], 0.685 with 95% CI [0.667, 0.699], and 0.505 with 95% CI [0.489, 0.539], respectively. Area Under the Precision Recall Curve was 0.684 with 95% CI [0.679, 0.702], 0.647 with 95% CI [0.638, 0.664], 0.429 with 95% CI [0.417, 0.434], and 0.551 with 95% CI [0.551, 0.552], respectively. Gradient Boosted Decision Trees was the best performing model of the supervised machine learning algorithms with satisfactory outcomes in the test cohort for predicting voided Computerized Physician Order Entry in Oral and Maxillofacial Surgery inpatients.

Online interpretable dynamic prediction models for clinically significant posthepatectomy liver failure based on machine learning algorithms: A retrospective cohort study.

Posthepatectomy liver failure (PHLF) is the leading cause of mortality in patients undergoing hepatectomy. However, practical models for accurately predicting the risk of PHLF are lacking. This study aimed to develop precise prediction models for clinically significant PHLF. A total of 226 patients undergoing hepatectomy at a single center were recruited. The study outcome was clinically significant PHLF. Five pre- and postoperative machine learning (ML) models were developed and compared with four clinical scores, namely, the MELD, FIB-4, ALBI, and APRI scores. The robustness of the developed ML models was internally validated using 5-fold cross-validation by calculating the average of the evaluation metrics and was externally validated on an independent temporal dataset, including the area under the curve (AUC) and the area under the precision‒recall curve (AUPRC). SHapley Additive exPlanations analysis was performed to interpret the best performance model. Clinically significant PHLF was observed in 23 of 226 patients (10.2%). The variables in the preoperative model included creatinine, total bilirubin, and Child‒Pugh grade. In addition to the above factors, the extent of resection was also a key variable for the postoperative model. The pre- and postoperative artificial neural network (ANN) models exhibited excellent performance, with mean AUCs of 0.766 and 0.851, respectively, and mean AUPRC values of 0.441 and 0.645, whereas the MELD, FIB-4, ALBI, and APRI scores reached AUCs of 0.714, 0.498, 0.536 and 0.551, respectively, and AUPRC values of 0.204, 0.111, 0.128 and 0.163, respectively. In addition, the AUCs of the pre- and postoperative ANN models were 0.720 and 0.731, respectively, and the AUPRC values were 0.380 and 0.408, respectively, on the temporal dataset. Our online interpretable dynamic ML models outperformed common clinical scores and could function as a clinical decision support tool to identify patients at high risk of PHLF pre- and postoperatively.

Deep learning based identification and interpretability research of traditional village heritage value elements: a case study in Hubei Province

The preservation and transmission of traditional villages is crucial to the prosperity and development of ethnic cultures. However, current traditional village surveys usually require a large number of experts and scholars to conduct field research, which is an expensive and time-consuming method, especially for large-scale tasks. Therefore, this study proposes an automatic classification method based on deep learning (DL) for the identification of traditional village heritage value elements (TVHVE). The study evaluates four selected convolutional neural network (CNN) frames using traditional villages in Hubei Province as a sample dataset. The results show that Residual Network152 (ResNet152) is the most suitable CNN frame for identifying TVHVE in Hubei. The stability and consistency of various TVHVE present in the ResNet152 model were evaluated using Area Under Curve (AUC) and Precision Recall Curve (PRC), which indicated satisfactory prediction performance for most elements, except for specific elements such as tombstones and stone carvings, which showed lower accuracy. In addition, the study sheds light on the areas of concern of the model with respect to different TVHVE images and elucidates the reasons behind the confusion between elements through semantic clustering based on image classification and interpretability analysis using the Gradient-Weighted Class Activation Mapping (Grad-CAM) heat map. By using an automated classification method based on DL, this study significantly reduces the cost and effort associated with traditional surveys. At the same time, insight into areas of concern and confusion in the model improves guidance for conservation efforts and provides valuable references for subsequent research.

Patient stratification based on the risk of severe illness in emergency departments through collaborative machine learning models

ObjectivesEmergency department (ED) overcrowding presents a global challenge that inhibits prompt care for critically ill patients. Traditional 5-level triage system that heavily rely on the judgment of the triage staff could fail to detect subtle symptoms in critical patients, thus leading to delayed treatment. Unlike previous rivalry-focused approaches, our study aimed to establish a collaborative machine learning (ML) model that renders risk scores for severe illness, which may assist the triage staff to provide a better patient stratification for timely critical cares. MethodsThis retrospective study was conducted at a tertiary teaching hospital. Data were collected from January 2015 to October 2022. Demographic and clinical information were collected at triage. The study focused on severe illness as the outcome. We developed artificial neural network (ANN) models, with or without utilizing the Taiwan Triage and Acuity Scale (TTAS) score as one of the predictors. The model using the TTAS score is termed a machine-human collaborative model (ANN-MH), while the model without it is referred to as a machine-only model (ANN-MO). The predictive power of these models was assessed using the area under the receiver-operating-characteristic (AUROC) and the precision-recall curves (AUPRC); their sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score were compared. ResultsThe study analyzed 668,602 ED visits from 2015 to 2022. Among them, 278,724 visits from 2015 to 2018 were used for model training and validation, while 320,201 visits from 2019 to 2022 were for testing model performance. Approximately 2.6% of visits were by severely ill patients, whose TTAS scores ranged from 1 to 5. The ANN-MH model achieved a testing AUROC of 0.918 and AUPRC of 0.369, while for the ANN-MO model the AUROC and AUPRC were 0.909 and 0.339, respectively. Based on these metrics, the ANN-MH model outperformed the ANN-MO model, and both surpassed human triage classification. Subgroup analyses further highlighted the models' capability to identify higher-risk patients within the same triage level. ConclusionsThe traditional 5-level triage system often falls short, leading to under-triage of critical patients. Our models include a score-based differentiation within a triage level to offer advanced risk stratification, thereby promoting patient safety.

The suboptimal clinical applicability of prognostic prediction models for severe postpartum hemorrhage: a meta-epidemiological study

ObjectivesTo systematically investigate clinical applicability of the current prognostic prediction models for severe postpartum hemorrhage (SPPH). Study Design and SettingA meta-epidemiological study of prognostic prediction models was conducted for SPPH. A pre-designed structured questionnaire was adopted to extract the study characteristics, predictors and the outcome, modeling methods, predictive performance, the classification ability for high-risk individuals, and clinical use scenarios. The risk of bias among studies was assessed by the Prediction model Risk Of Bias ASsessment Tool (PROBAST). ResultsTwenty-two studies containing 27 prediction models were included. The number of predictors in the final models varied from 3 to 53. However, one-third of the models (11) did not clearly specify the timing of predictor measurement. Calibration was found to be lacking in 10 (37.0%) models. Among the 20 models with an incidence rate of predicted outcomes below 15.0%, none of the models estimated the area under the precision-recall curve, and all reported positive predictive values were below 40.0%. Only two (7.4%) models specified the target clinical setting, while seven (25.9%) models clarified the intended timing of model use. Lastly, all 22 studies were deemed to be at high risk of bias. ConclusionCurrent SPPH prediction models have limited clinical applicability due to methodological flaws, including unclear predictor measurement, inadequate calibration assessment, and insufficient evaluation of classification ability. Additionally, there is a lack of clarity regarding the timing for model use, target users, and clinical settings. These limitations raise concerns about the reliability and usefulness of these models in real-world clinical practice.

A deep learning-based automated diagnosis system for SPECT myocardial perfusion imaging

Images obtained from single-photon emission computed tomography for myocardial perfusion imaging (MPI SPECT) contain noises and artifacts, making cardiovascular disease diagnosis difficult. We developed a deep learning-based diagnosis support system using MPI SPECT images. Single-center datasets of MPI SPECT images (n = 5443) were obtained and labeled as healthy or coronary artery disease based on diagnosis reports. Three axes of four-dimensional datasets, resting, and stress conditions of three-dimensional reconstruction data, were reconstructed, and an AI model was trained to classify them. The trained convolutional neural network showed high performance [area under the curve (AUC) of the ROC curve: approximately 0.91; area under the recall precision curve: 0.87]. Additionally, using unsupervised learning and the Grad-CAM method, diseased lesions were successfully visualized. The AI-based automated diagnosis system had the highest performance (88%), followed by cardiologists with AI-guided diagnosis (80%) and cardiologists alone (65%). Furthermore, diagnosis time was shorter for AI-guided diagnosis (12 min) than for cardiologists alone (31 min). Our high-quality deep learning-based diagnosis support system may benefit cardiologists by improving diagnostic accuracy and reducing working hours.

Class imbalance on medical image classification: towards better evaluation practices for discrimination and calibration performance.

This work aims to assess standard evaluation practices used by the research community for evaluating medical imaging classifiers, with a specific focus on the implications of class imbalance. The analysis is performed on chest X-rays as a case study and encompasses a comprehensive model performance definition, considering both discriminative capabilities and model calibration. We conduct a concise literature review to examine prevailing scientific practices used when evaluating X-ray classifiers. Then, we perform a systematic experiment on two major chest X-ray datasets to showcase a didactic example of the behavior of several performance metrics under different class ratios and highlight how widely adopted metrics can conceal performance in the minority class. Our literature study confirms that: (1) even when dealing with highly imbalanced datasets, the community tends to use metrics that are dominated by the majority class; and (2) it is still uncommon to include calibration studies for chest X-ray classifiers, albeit its importance in the context of healthcare. Moreover, our systematic experiments confirm that current evaluation practices may not reflect model performance in real clinical scenarios and suggest complementary metrics to better reflect the performance of the system in such scenarios. Our analysis underscores the need for enhanced evaluation practices, particularly in the context of class-imbalanced chest X-ray classifiers. We recommend the inclusion of complementary metrics such as the area under the precision-recall curve (AUC-PR), adjusted AUC-PR, and balanced Brier score, to offer a more accurate depiction of system performance in real clinical scenarios, considering metrics that reflect both, discrimination and calibration performance. This study underscores the critical need for refined evaluation metrics in medical imaging classifiers, emphasizing that prevalent metrics may mask poor performance in minority classes, potentially impacting clinical diagnoses and healthcare outcomes. Common scientific practices in papers dealing with X-ray computer-assisted diagnosis (CAD) systems may be misleading. We highlight limitations in reporting of evaluation metrics for X-ray CAD systems in highly imbalanced scenarios. We propose adopting alternative metrics based on experimental evaluation on large-scale datasets.

Development and validation of a LASSO prediction model for cisplatin induced nephrotoxicity: a case-control study in China

BackgroundEarly identification of high-risk individuals with cisplatin-induced nephrotoxicity (CIN) is crucial for avoiding CIN and improving prognosis. In this study, we developed and validated a CIN prediction model based on general clinical data, laboratory indications, and genetic features of lung cancer patients before chemotherapy.MethodsWe retrospectively included 696 lung cancer patients using platinum chemotherapy regimens from June 2019 to June 2021 as the traing set to construct a predictive model using Absolute shrinkage and selection operator (LASSO) regression, cross validation, and Akaike’s information criterion (AIC) to select important variables. We prospectively selected 283 independent lung cancer patients from July 2021 to December 2022 as the test set to evaluate the model’s performance.ResultsThe prediction model showed good discrimination and calibration, with AUCs of 0.9217 and 0.8288, sensitivity of 79.89% and 45.07%, specificity of 94.48% and 94.81%, in the training and test sets respectively. Clinical decision curve analysis suggested that the model has value for clinical use when the risk threshold ranges between 0.1 and 0.9. Precision-Recall (PR) curve shown in recall interval from 0.5 to 0.75: precision gradually declines with increasing Recall, up to 0.9.ConclusionsPredictive models based on laboratory and demographic variables can serve as a beneficial complementary tool for identifying high-risk populations with CIN.

D'or: deep orienter of protein-protein interaction networks.

Protein-protein interactions (PPIs) provide the skeleton for signal transduction in the cell. Current PPI measurement techniques do not provide information on their directionality which is critical for elucidating signaling pathways. To date, there are hundreds of thousands of known PPIs in public databases, yet only a small fraction of them have an assigned direction. This information gap calls for computational approaches for inferring the directionality of PPIs, aka network orientation. In this work, we propose a novel deep learning approach for PPI network orientation. Our method first generates a set of proximity scores between a protein interaction and sets of cause and effect proteins using a network propagation procedure. Each of these score sets is fed, one at a time, to a deep set encoder whose outputs are used as features for predicting the interaction's orientation. On a comprehensive dataset of oriented PPIs taken from five different sources, we achieve an area under the precision-recall curve of 0.89-0.92, outperforming previous methods. We further demonstrate the utility of the oriented network in prioritizing cancer driver genes and disease genes. D'or is implemented in Python and is publicly available at https://github.com/pirakd/DeepOrienter.

Use of an Artificial Intelligence-Generated Vascular Severity Score Improved Plus Disease Diagnosis in Retinopathy of Prematurity

To evaluate whether providing clinicians with an artificial intelligence (AI)-based vascular severity score (VSS) improves consistency in the diagnosis of plus disease in retinopathy of prematurity (ROP). Multireader diagnostic accuracy imaging study. Eleven ROP experts, 9 of whom had been in practice for 10 years or more. RetCam (Natus Medical Incorporated) fundus images were obtained from premature infants during routine ROP screening as part of the Imaging and Informatics in ROP study between January 2012 and July 2020. From all available examinations, a subset of 150 eye examinations from 110 infants were selected for grading. An AI-based VSS was assigned to each set of images using the i-ROP DL system (Siloam Vision). The clinicians were asked to diagnose plus disease for each examination and to assign an estimated VSS (range, 1-9) at baseline, and then again 1 month later with AI-based VSS assistance. A reference standard diagnosis (RSD) was assigned to each eye examination from the Imaging and Informatics in ROP study based on 3 masked expert labels and the ophthalmoscopic diagnosis. Mean linearly weighted κ value for plus disease diagnosis compared with RSD. Area under the receiver operating characteristic curve (AUC) and area under the precision-recall curve (AUPR) for labels 1 through 9 compared with RSD for plus disease. Expert agreement improved significantly, from substantial (κ value, 0.69 [0.59, 0.75]) to near perfect (κ value, 0.81 [0.71, 0.86]), when AI-based VSS was integrated. Additionally, a significant improvement in plus disease discrimination was achieved as measured by mean AUC (from 0.94 [95% confidence interval (CI), 0.92-0.96] to 0.98 [95% CI, 0.96-0.99]; difference, 0.04 [95% CI, 0.01-0.06]) and AUPR (from 0.86 [95% CI, 0.81-0.90] to 0.95 [95% CI, 0.91-0.97]; difference, 0.09 [95% CI, 0.03-0.14]). Providing ROP clinicians with an AI-based measurement of vascular severity in ROP was associated with both improved plus disease diagnosis and improved continuous severity labeling as compared with an RSD for plus disease. If implemented in practice, AI-based VSS could reduce interobserver variability and could standardize treatment for infants with ROP. Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Multi-sequence MRI-based radiomics: An objective method to diagnose early-stage osteonecrosis of the femoral head

ObjectivesThis study investigated the use of radiomics for diagnosing early-stage osteonecrosis of the femoral head (ONFH) by extracting features from multiple MRI sequences and constructing predictive models. Materials and methodsWe conducted a retrospective review, collected MR images of early-stage ONFH (102 from institution A and 20 from institution B) and healthy femoral heads (102 from institution A and 20 from institution B) from two institutions. We extracted radiomics features, handled batch effects using Combat, and normalized features using z-score. We employed the Least absolute shrinkage and selection operator (LASSO) algorithm, along with Max-Relevance and Min-Redundancy (mRMR), to select optimal features for constructing radiomics models based on single, double, and multi-sequence MRI data. We evaluated performance using receiver operating characteristic (ROC) and precision-recall (PR) curves, and compared area under curve of ROC (AUC-ROC) values with the DeLong test. Additionally, we studied the diagnostic performance of the multi-sequence radiomics model and radiologists, compared the diagnostic outcomes of the model and radiologists using the Fisher exact test. ResultsWe studied 122 early-stage ONFH and 122 normal femoral heads. The multi-sequence model exhibited the best diagnostic performance among all models (AUC-ROC, PR-AUC for training set: 0.96, 0.961; validation set: 0.96, 0.97; test set: 0.94, 0.94), and it outperformed three resident radiologists on the external testing group with an accuracy of 87.5 %, sensitivity of 85.00 %, and specificity of 90.00 % (p < 0.01), highlighting the robustness of our findings. ConclusionsOur study underscored the novelty of the multi-sequence radiomics model in diagnosing early-stage ONFH. By leveraging features extracted from multiple imaging sequences, this approach demonstrated high efficacy, indicating its potential to advance early diagnosis for ONFH. These findings provided important guidance for enhancing early diagnosis of ONFH through radiomics methods, offering new avenues and possibilities for clinical practice and patient care.

Self-supervised learning for classifying paranasal anomalies in the maxillary sinus

PurposeParanasal anomalies, frequently identified in routine radiological screenings, exhibit diverse morphological characteristics. Due to the diversity of anomalies, supervised learning methods require large labelled dataset exhibiting diverse anomaly morphology. Self-supervised learning (SSL) can be used to learn representations from unlabelled data. However, there are no SSL methods designed for the downstream task of classifying paranasal anomalies in the maxillary sinus (MS).MethodsOur approach uses a 3D convolutional autoencoder (CAE) trained in an unsupervised anomaly detection (UAD) framework. Initially, we train the 3D CAE to reduce reconstruction errors when reconstructing normal maxillary sinus (MS) image. Then, this CAE is applied to an unlabelled dataset to generate coarse anomaly locations by creating residual MS images. Following this, a 3D convolutional neural network (CNN) reconstructs these residual images, which forms our SSL task. Lastly, we fine-tune the encoder part of the 3D CNN on a labelled dataset of normal and anomalous MS images.ResultsThe proposed SSL technique exhibits superior performance compared to existing generic self-supervised methods, especially in scenarios with limited annotated data. When trained on just 10% of the annotated dataset, our method achieves an area under the precision-recall curve (AUPRC) of 0.79 for the downstream classification task. This performance surpasses other methods, with BYOL attaining an AUPRC of 0.75, SimSiam at 0.74, SimCLR at 0.73 and masked autoencoding using SparK at 0.75.ConclusionA self-supervised learning approach that inherently focuses on localizing paranasal anomalies proves to be advantageous, particularly when the subsequent task involves differentiating normal from anomalous maxillary sinuses. Access our code at https://github.com/mtec-tuhh/self-supervised-paranasal-anomaly.

Predictive approach for liberation from acute dialysis in ICU patients using interpretable machine learning

Renal recovery following dialysis-requiring acute kidney injury (AKI-D) is a vital clinical outcome in critical care, yet it remains an understudied area. This retrospective cohort study, conducted in a medical center in Taiwan from 2015 to 2020, enrolled patients with AKI-D during intensive care unit stays. We aimed to develop and temporally test models for predicting dialysis liberation before hospital discharge using machine learning algorithms and explore early predictors. The dataset comprised 90 routinely collected variables within the first three days of dialysis initiation. Out of 1,381 patients who received acute dialysis, 27.3% experienced renal recovery. The cohort was divided into the training group (N = 1135) and temporal testing group (N = 251). The models demonstrated good performance, with an area under the receiver operating characteristic curve of 0.85 (95% CI, 0.81–0.88) and an area under the precision-recall curve of 0.69 (95% CI, 0.62–0.76) for the XGBoost model. Key predictors included urine volume, Charlson comorbidity index, vital sign derivatives (trend of respiratory rate and SpO2), and lactate levels. We successfully developed early prediction models for renal recovery by integrating early changes in vital signs and inputs/outputs, which have the potential to aid clinical decision-making in the ICU.

Applying machine learning to Galactic Archaeology: how well can we recover the origin of stars in Milky Way-like galaxies?

ABSTRACT We present several machine learning (ML) models developed to efficiently separate stars formed in situ in Milky Way-type galaxies from those that were formed externally and later accreted. These models, which include examples from artificial neural networks, decision trees, and dimensionality reduction techniques, are trained on a sample of disc-like, Milky Way-mass galaxies drawn from the artemis cosmological hydrodynamical zoom-in simulations. We find that the input parameters which provide an optimal performance for these models consist of a combination of stellar positions, kinematics, chemical abundances ([Fe/H] and [α/Fe]), and photometric properties. Models from all categories perform similarly well, with area under the precision–recall curve (PR-AUC) scores of ≃ 0.6. Beyond a galactocentric radius of 5 kpc, models retrieve $\gt 90~{{\ \rm per\ cent}}$ of accreted stars, with a sample purity close to 60 per cent, however the purity can be increased by adjusting the classification threshold. For one model, we also include host galaxy-specific properties in the training, to account for the variability of accretion histories of the hosts, however this does not lead to an improvement in performance. The ML models can identify accreted stars even in regions heavily dominated by the in-situ component (e.g. in the disc), and perform well on an unseen suite of simulations (the auriga simulations). The general applicability bodes well for application of such methods on observational data to identify accreted substructures in the Milky Way without the need to resort to selection cuts for minimizing the contamination from in-situ stars.

Improving accuracy of code smells detection using machine learning with data balancing techniques

Code smells indicate potential symptoms or problems in software due to inefficient design or incomplete implementation. These problems can affect software quality in the long-term. Code smell detection is fundamental to improving software quality and maintainability, reducing software failure risk, and helping to refactor the code. Previous works have applied several prediction methods for code smell detection. However, many of them show that machine learning (ML) and deep learning (DL) techniques are not always suitable for code smell detection due to the problem of imbalanced data. So, data imbalance is the main challenge for ML and DL techniques in detecting code smells. To overcome these challenges, this study aims to present a method for detecting code smell based on DL algorithms (Bidirectional Long Short-Term Memory (Bi-LSTM) and Gated Recurrent Unit (GRU)) combined with data balancing techniques (random oversampling and Tomek links) to mitigate data imbalance issue. To establish the effectiveness of the proposed models, the experiments were conducted on four code smells datasets (God class, data Class, feature envy, and long method) extracted from 74 open-source systems. We compare and evaluate the performance of the models according to seven different performance measures accuracy, precision, recall, f-measure, Matthew’s correlation coefficient (MCC), the area under a receiver operating characteristic curve (AUC), the area under the precision–recall curve (AUCPR) and mean square error (MSE). After comparing the results obtained by the proposed models on the original and balanced data sets, we found out that the best accuracy of 98% was obtained for the Long method by using both models (Bi-LSTM and GRU) on the original datasets, the best accuracy of 100% was obtained for the long method by using both models (Bi-LSTM and GRU) on the balanced datasets (using random oversampling), and the best accuracy 99% was obtained for the long method by using Bi-LSTM model and 99% was obtained for the data class and Feature envy by using GRU model on the balanced datasets (using Tomek links). The results indicate that the use of data balancing techniques had a positive effect on the predictive accuracy of the models presented. The results show that the proposed models can detect the code smells more accurately and effectively.