External Validation Dataset Research Articles

Automatic deep learning method for third lumbar selection and body composition evaluation on CT scans of cancer patients.

The importance of body composition and sarcopenia is well-recognized in cancer patient outcomes and treatment tolerance, yet routine evaluations are rare due to their time-intensive nature. While CT scans provide accurate measurements, they depend on manual processes. We developed and validated a deep learning algorithm to automatically select and segment abdominal muscles [SM], visceral fat [VAT], and subcutaneous fat [SAT] on CT scans. A total of 352 CT scans were collected from two cancer centers. The detection of the third lumbar vertebra and three different body tissues (SM, VAT, and SAT) were annotated manually. The 5-fold cross-validation method was used to develop the algorithm and validate its performance on the training cohort. The results were validated on an external, independent group of CT scans. The algorithm for automatic L3 slice selection had a mean absolute error of 4 mm for the internal validation dataset and 5.5 mm for the external validation dataset. The median DICE similarity coefficient for body composition was 0.94 for SM, 0.93 for VAT, and 0.86 for SAT in the internal validation dataset, whereas it was 0.93 for SM, 0.93 for VAT, and 0.85 for SAT in the external validation dataset. There were high correlation scores with sarcopenia metrics in both internal and external validation datasets. Our deep learning algorithm facilitates routine research use and could be integrated into electronic patient records, enhancing care through better monitoring and the incorporation of targeted supportive measures like exercise and nutrition.

Deep embedded clustering generalisability and adaptation for integrating mixed datatypes: two critical care cohorts

We validated a Deep Embedded Clustering (DEC) model and its adaptation for integrating mixed datatypes (in this study, numerical and categorical variables). Deep Embedded Clustering (DEC) is a promising technique capable of managing extensive sets of variables and non-linear relationships. Nevertheless, DEC cannot adequately handle mixed datatypes. Therefore, we adapted DEC by replacing the autoencoder with an X-shaped variational autoencoder (XVAE) and optimising hyperparameters for cluster stability. We call this model “X-DEC”. We compared DEC and X-DEC by reproducing a previous study that used DEC to identify clusters in a population of intensive care patients. We assessed internal validity based on cluster stability on the development dataset. Since generalisability of clustering models has insufficiently been validated on external populations, we assessed external validity by investigating cluster generalisability onto an external validation dataset. We concluded that both DEC and X-DEC resulted in clinically recognisable and generalisable clusters, but X-DEC produced much more stable clusters.

A large-scale study of peptide features defining immunogenicity of cancer neo-epitopes.

Accurate prediction of immunogenicity for neo-epitopes arising from a cancer associated mutation is a crucial step in many bioinformatics pipelines that predict outcome of checkpoint blockade treatments or that aim to design personalised cancer immunotherapies and vaccines. In this study, we performed a comprehensive analysis of peptide features relevant for prediction of immunogenicity using the Cancer Epitope Database and Analysis Resource (CEDAR), a curated database of cancer epitopes with experimentally validated immunogenicity annotations from peer-reviewed publications. The developed model, ICERFIRE (ICore-based Ensemble Random Forest for neo-epitope Immunogenicity pREdiction), extracts the predicted ICORE from the full neo-epitope as input, i.e. the nested peptide with the highest predicted major histocompatibility complex (MHC) binding potential combined with its predicted likelihood of antigen presentation (%Rank). Key additional features integrated into the model include assessment of the BLOSUM mutation score of the neo-epitope, and antigen expression levels of the wild-type counterpart which is often reflecting a neo-epitope's abundance. We demonstrate improved and robust performance of ICERFIRE over existing immunogenicity and epitope prediction models, both in cross-validation and on external validation datasets.

Deep learning–radiomics integrated noninvasive detection of epidermal growth factor receptor mutations in non-small cell lung cancer patients

This study focused on a novel strategy that combines deep learning and radiomics to predict epidermal growth factor receptor (EGFR) mutations in patients with non-small cell lung cancer (NSCLC) using computed tomography (CT). A total of 1280 patients with NSCLC who underwent contrast-enhanced CT scans and EGFR mutation testing before treatment were selected for the final study. Regions of interest were segmented from the CT images to extract radiomics features and obtain tumor images. These tumor images were input into a convolutional neural network model to extract 512 image features, which were combined with radiographic features and clinical data to predict the EGFR mutation. The generalization performance of the model was evaluated using external institutional data. The internal and external datasets contained 324 and 130 EGFR mutants, respectively. Sex, height, weight, smoking history, and clinical stage were significantly different between the EGFR-mutant patient groups. The EGFR mutations were predicted by combining the radiomics and clinical features, and an external validation dataset yielded an area under the curve (AUC) value of 0.7038. The model utilized 1280 tumor images, radiomics features, and clinical characteristics as input data and exhibited an AUC of approximately 0.81 and 0.78 during the primary cohort and external validation, respectively. These results indicate the feasibility of integrating radiomics analysis with deep learning for predicting EGFR mutations. CT-image-based genetic testing is a simple EGFR mutation prediction method, which can improve the prognosis of NSCLC patients and help establish personalized treatment strategies.

Nomogram based on high-density lipoprotein cholesterol for the occurrence of preoperative deep vein thrombosis in patients with intertrochanteric femur fracture: a retrospective study

BackgroundThis study aims to develop a nomogram and forecast the incidence of DVT in individuals suffering from an intertrochanteric femur fracture.MethodThis work created a nomogram using the R programming language and employed logistic regression to determine independent predicting features. An external validation dataset was used to validate the nomogram.ResultThe findings demonstrated the independence of LYM (0.02[0.01–0.09], p < 0.001), ALB (0.83[0.74, 0.94], p = 0.002), and HDL-C (0.18[0.04, 0.71], p = 0.014). Good prediction performance with modest errors was shown by the nomogram in both the training and validation groups.ConclusionIn conclusion, the nomogram that was created using HDL-C, ALB, and LYM can assist medical professionals in determining the likelihood that DVT will occur.

Identification and characterization of biomarkers associated with endoplasmic reticulum protein processing in cerebral ischemia-reperfusion injury.

Cerebral ischemia (CI), ranking as the second leading global cause of death, is frequently treated by reestablishing blood flow and oxygenation. Paradoxically, this reperfusion can intensify tissue damage, leading to CI-reperfusion injury. This research sought to uncover biomarkers pertaining to protein processing in the endoplasmic reticulum (PER) during CI-reperfusion injury. We utilized the Gene Expression Omnibus (GEO) dataset GSE163614 to discern differentially expressed genes (DEGs) and single out PER-related DEGs. The functions and pathways of these PER-related DEGs were identified via Gene Ontology (GO) and the Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses. Core genes were pinpointed through protein-protein interaction (PPI) networks. Subsequent to this, genes with diagnostic relevance were distinguished using external validation datasets. A single-sample gene-set enrichment analysis (ssGSEA) was undertaken to pinpoint genes with strong associations to hypoxia and apoptosis, suggesting their potential roles as primary inducers of apoptosis in hypoxic conditions during ischemia-reperfusion injuries. Our study demonstrated that PER-related genes, specifically ADCY5, CAMK2A, PLCB1, NTRK2, and DLG4, were markedly down-regulated in models, exhibiting a robust association with hypoxia and apoptosis. The data indicates that ADCY5, CAMK2A, PLCB1, NTRK2, and DLG4 could be pivotal genes responsible for triggering apoptosis in hypoxic environments during CI-reperfusion injury.

Proxy measures for the assessment of psychotic and affective symptoms in studies using electronic health records.

There is a lack of standardised psychometric data in electronic health record (EHR)-based research. Proxy measures of symptom severity based on patients' clinical records may be useful surrogates in mental health EHR research. This study aimed to validate proxy tools for the short versions of the Positive and Negative Syndrome Scale (PANSS-6), Young Mania Rating Scale (YMRS-6) and Montgomery-Åsberg Depression Rating Scale (MADRS-6). A cross-sectional, multicentre study was conducted in a sample of 116 patients with first-episode psychosis from 12 public hospitals in Spain. Concordance between PANSS-6, YMRS-6 and MADRS-6 scores and their respective proxies was evaluated based on information from EHR clinical notes, using a variety of statistical procedures, including multivariate tests to adjust for potential confounders. Bootstrapping techniques were used for internal validation, and an independent cohort from the Treatment and Early Intervention in Psychosis Program (TIPP-Lausanne, Switzerland) for external validation. The proxy versions correlated strongly with their respective standardised scales (partial correlations ranged from 0.75 to 0.84) and had good accuracy and discriminatory power in distinguishing between patients in and not in remission (percentage of patients correctly classified ranged from 83.9 to 91.4% and bootstrapped optimism-corrected area under the receiver operating characteristic curve ranged from 0.76 to 0.89), with high interrater reliability (intraclass correlation coefficient of 0.81). The findings remained robust in the external validation data-set. The proxy instruments proposed for assessing psychotic and affective symptoms by reviewing EHR provide a feasible and reliable alternative to traditional structured psychometric procedures, and a promising methodology for real-world practice settings.

Exploration of ferroptosis and necroptosis-related genes and potential molecular mechanisms in psoriasis and atherosclerosis.

Ferroptosis and necroptosis are two recently identified forms of non-apoptotic cell death. Their dysregulation plays a critical role in the development and progression of Psoriasis (PsD) and Atherosclerosis (AS). This study explores shared Ferroptosis and necroptosis-related genes and elucidates their molecular mechanisms in PsD and AS through the analysis of public databases. Data sets for PsD (GSE30999) and AS (GSE28829) were retrieved from the GEO database. Differential gene expression (DEG) and weighted gene co-expression network analysis (WGCNA) were performed. Machine learning algorithms identified candidate biomarkers, whose diagnostic values were assessed using Receiver Operating Characteristic (ROC) curve analysis. Additionally, the expression levels of these biomarkers in cell models of AS and PsD were quantitatively measured using Western Blot (WB) and real-time quantitative PCR (RT-qPCR). Furthermore, CIBERSORT evaluated immune cell infiltration in PsD and AS tissues, highlighting the correlation between characteristic genes and immune cells. Predictive analysis for candidate drugs targeting characteristic genes was conducted using the DGIdb database, and an lncRNA-miRNA-mRNA network related to these genes was constructed. We identified 44 differentially expressed ferroptosis-related genes (DE-FRGs) and 30 differentially expressed necroptosis-related genes (DE-NRGs). GO and KEGG enrichment analyses revealed significant enrichment of these genes in immune-related and inflammatory pathways, especially in NOD-like receptor and TNF signaling pathways. Two ferroptosis-related genes (NAMPT, ZFP36) and eight necroptosis-related genes (C7, CARD6, CASP1, CTSD, HMOX1, NOD2, PYCARD, TNFRSF21) showed high sensitivity and specificity in ROC curve analysis. These findings were corroborated in external validation datasets and cell models. Immune infiltration analysis revealed increased levels of T cells gamma delta, Macrophages M0, and Macrophages M2 in PsD and AS samples. Additionally, we identified 43 drugs targeting 5 characteristic genes. Notably, the XIST-miR-93-5p-ZFP36/HMOX1 and NEAT1-miR-93-5p-ZFP36/HMOX1 pathways have been identified as promising RNA regulatory pathways in AS and PsD. The two ferroptosis-related genes (NAMPT, ZFP36) and eight necroptosis-related genes (C7, CARD6, CASP1, CTSD, HMOX1, NOD2, PYCARD, TNFRSF21) are potential key biomarkers for PsD and AS. These genes significantly influence the pathogenesis of PsD and AS by modulating macrophage activity, participating in immune regulation, and mediating inflammatory responses.

A practical nomogram for preoperatively predicting lateral cervical lymph node metastasis in medullary thyroid carcinoma: a dual-center retrospective study.

Lateral lymph node metastasis (LLNM) is very common in medullary thyroid carcinoma (MTC), but there is still controversy about how to manage cervical lateral lymph nodes, especially for clinically negative MTC. The aim of this study is to develop and validate a nomogram for predicting LLNM risk in MTC. A total of 234 patients from two hospitals were retrospectively enrolled in this study and divided into LLNM positive group and LLNM negative group based on the pathology. The correlation between LLNM and preoperative clinical and ultrasound variables were evaluated by univariable and multivariable logistic regression analysis. A nomogram was generated to predict the risk of the LLNM of MTC patients, validated by external dataset, and evaluated in terms of discrimination, calibration, and clinical usefulness. The training, internal, and external validation datasets included 152, 51, and 31 MTC patients, respectively. According to the multivariable logistic regression analysis, gender (male), relationship to thyroid capsule and serum calcitonin were independently associated with LLNM in the training dataset. The predictive nomogram model developed with the aforementioned variables showed favorable performance in estimating risk of LLNM, with the area under the ROC curve (AUC) of 0.826 in the training dataset, 0.816 in the internal validation dataset, and 0.846 in the external validation dataset. We developed and validated a model named MTC nomogram, utilizing available preoperative variables to predict the probability of LLNM in patients with MTC. This nomogram will be of great value for guiding the clinical diagnosis and treatment process of MTC patients.

A comprehensive evaluation framework for benchmarking multi-objective feature selection in omics-based biomarker discovery.

Machine learning algorithms have been extensively used for accurate classification of cancer subtypes driven by gene expression-based biomarkers. However, biomarker models combining multiple gene expression signatures are often not reproducible in external validation datasets and their feature set size is often not optimized, jeopardizing their translatability into cost-effective clinical tools. We investigated how to solve the multi-objective problem of finding the best trade-offs between classification performance and set size applying seven algorithms for machine learning-driven feature subset selection and analyse how they perform in a benchmark with eight large-scale transcriptome datasets of cancer, covering both training and external validation sets. The benchmark includes evaluation metrics assessing the performance of the individual biomarkers and the solution sets, according to their accuracy, diversity, and stability of the composing genes. Moreover, a new evaluation metric for cross-validation studies is proposed that generalizes the hypervolume, which is commonly used to assess the performance of multi-objective optimization algorithms. Biomarkers exhibiting 0.8 of balanced accuracy on the external dataset for breast, kidney and ovarian cancer using respectively 4, 2 and 7 features, were obtained. Genetic algorithms often provided better performance than other considered algorithms, and the recently proposed NSGA2-CH and NSGA2-CHS were the best performing methods in most cases.

Ultrasound Radiomics for the Prediction of Breast Cancers with HER2-Zero, -Low, and -Positive Status: A Dual-Center Study.

To assess whether gray-scale ultrasound (US) based radiomic features can help distinguish HER2 expressions (ie, HER2-overexpressing, HER2-low-expressing, and HER2-zero-expressing) in breast cancer. This retrospective study encompassed female breast cancer patients who underwent US examinations at two distinct centers from February 2021 to July 2023. Tumor segmentation and radiomic feature extraction were performed on grayscale US images. Decision Tree analysis was employed to simultaneously evaluate feature importance, and the Least Absolute Shrinkage and Selection Operator technique was utilized for feature selection to construct the radiomic signature. The Area Under the Curve (AUC) of the Receiver Operating Characteristic curve was employed to assess the performance of the radiomic features. Multivariate logistic regression was used to identify independent predictors for distinguishing HER2 expression in the dataset. The training set comprised 292 patients from Center 1 (median, 51 years; interquartile range [IQR]: 45-61), while the external validation set included 131 patients from Center 2 (median, 51 years; IQR: 45-62). In the external validation dataset, the radiomic features achieved AUC of 0.76 for distinguishing between HER2-low and positive tumors versus HER2-zero tumors. The AUC for differentiating HER2-low (1+) from HER2-zero tumors was 0.74, and for distinguishing HER2-low (2+) from HER2-zero tumors, the AUC was 0.77. In the multivariate analysis assessing HER2-low and HER2-positive versus HER2-zero tumors, internal echoes (P = .029) and margins (P < .001) emerged as independent predictive factors. The radiomic signature and tumor descriptors from gray-scale US may predict distinct HER2 expressions of breast cancers with therapeutic implications.

Construction and comparison of short-term prognosis prediction model based on machine learning in acute ischemic stroke

ObjectiveTo construct and compared the short-term prognosis prediction models of acute ischemic stroke (AIS) by machine learning (ML). MethodsRetrospectively study. The group W (mRS≤3) was clustered, and combined with group P (mRS>3) to form the post-clustering dataset for modeling. The “glmnet”, “rpart”, “xgboost”, “randomForest”, “neuralnet” packages were used to construct ML models. The accuracy, sensitivity, specificity, positive predict value (PPV), negative predict value (NPV) among the models were compared. Four external clinical datasets were used for external clinical validation. The optimal prediction model was determined by variable screening ability, model visualization, and external clinical validation performance. ResultsThe post-clustering dataset contains 139 patients (group W) and 122 patients (group P). The neutrophil multiplied by D-dimer (NDM) has predictive value in all ML prediction models in this study. In the decision tree model, NDMQ occupies the first tree node, When NDM≤5.62 and the age<74.5, the probability of poor prognosis of AIS is less than 20 %. When NDM>5.62 and accompanied by pneumonia, the incidence of poor prognosis of AIS is about 90 %. In the Random Forest (RF) model, NDMQ had the highest Gini index. The variable combination screened by the RF model had the best performance in the neural network, and the accuracy, sensitivity, specificity, PPV, and NPV of the external validation were 0.800, 0.774, 0.833, 0.857, and 0.741, respectively. The RF model had the best performance in the external clinical validation datasets, with accuracies of 0.646, 0.697, 0.695, and 0.713, respectively. ConclusionsNDM shows predictive value for AIS short-term prognosis in all ML models in this study. The optimal model in screening characteristic variables and the performance of in external clinical datasets was RF model. In the analysis of medical data with small sample size and outcome as categorical variables, RF could be used as the main algorithm to build a model.

Predicting incident heart failure from population-based nationwide electronic health records: protocol for a model development and validation study

IntroductionHeart failure (HF) is increasingly common and associated with excess morbidity, mortality, and healthcare costs. Treatment of HF can alter the disease trajectory and reduce clinical events in HF. However,...

A Multi-Task Transformer With Local-Global Feature Interaction and Multiple Tumoral Region Guidance for Breast Cancer Diagnosis.

Breast cancer, as a malignant tumor disease, has maintained high incidence and mortality rates over the years. Ultrasonography is one of the primary methods for diagnosing early-stage breast cancer. However, correctly interpreting breast ultrasound images requires massive time from physicians with specialized knowledge and extensive experience. Recently, deep learning-based method have made significant advancements in breast tumor segmentation and classification due to their powerful fitting capabilities. However, most existing methods focus on performing one of these tasks separately, and often failing to effectively leverage information from specific tumor-related areas that hold considerable diagnostic value. In this study, we propose a multi-task network with local-global feature interaction and multiple tumoral region guidance for breast ultrasound-based tumor segmentation and classification. Specifically, we construct a dual-stream encoder, paralleling CNN and Transformer, to facilitate hierarchical interaction and fusion of local and global features. This architecture enables each stream to capitalize on the strengths of the other while preserving its unique characteristics. Moreover, we design a multi-tumoral region guidance module to explicitly learn long-range non-local dependencies within intra-tumoral and peri-tumoral regions from spatial domain, thus providing interpretable cues beneficial for classification. Experimental results on two breast ultrasound datasets show that our network outperforms state-of-the-art methods in tumor segmentation and classification tasks. Compared with the second-best competitive method, our network improves the diagnosis accuracy from 73.64% to 80.21% on a large external validation dataset, which demonstrates its superior generalization capability.

Artificial Intelligence Model for a Distinction between Early-Stage Gastric Cancer Invasive Depth T1a and T1b.

Background: Endoscopic submucosal dissection (ESD) is a widely accepted treatment for patients with mucosa (T1a) disease without lymph node metastasis. However, the inconsistency of inspection quality of tumor staging under the standard tool combining endoscopic ultrasound (EUS) with computed tomography (CT) scanning makes it restrictive. Methods: We conducted a study using data augmentation and artificial intelligence (AI) to address the early gastric cancer (EGC) staging problem. The proposed AI model simplifies early cancer treatment by eliminating the need for ultrasound or other staging methods. We developed an AI model utilizing data augmentation and the You-Only-Look-Once (YOLO) approach. We collected a white-light image dataset of 351 stage T1a and 542 T1b images to build, test, and validate the model. An external white-light images dataset that consists of 47 T1a and 9 T1b images was then collected to validate our AI model. The result of the external dataset validation indicated that our model also applies to other peer health institutes. Results: The results of k-fold cross-validation using the original dataset demonstrated that the proposed model had a sensitivity of 85.08% and an average specificity of 87.17%. Additionally, the k-fold cross-validation model had an average accuracy rate of 86.18%; the external data set demonstrated similar validation results with a sensitivity of 82.98%, a specificity of 77.78%, and an overall accuracy of 82.14%. Conclusions: Our findings suggest that the AI model can effectively replace EUS and CT in early GC staging, with an average validation accuracy rate of 86.18% for the original dataset from Linkou Cheng Gun Memorial Hospital and 82.14% for the external validation dataset from Kaohsiung Cheng Gun Memorial Hospital. Moreover, our AI model's accuracy rate outperformed the average EUS and CT rates in previous literature (around 70%).

IntraCranial pressure prediction AlgoRithm using machinE learning (I-CARE): Training and Validation Study.

Elevated intracranial pressure (ICP) is a potentially devastating complication of neurologic injury. Developing an ICP prediction algorithm to help the clinician adjust treatments and potentially prevent elevated ICP episodes. Retrospective study. Three hundred thirty-five ICUs at 208 hospitals in the United States. Adults patients from the electronic ICU (eICU) Collaborative Research Database was used to train an ensemble machine learning model to predict the ICP 30 minutes in the future. Predictive performance was evaluated using a left-out test dataset and externally evaluated on the Medical Information Mart for Intensive Care-III (MIMIC-III) Matched Waveform Database. None. Predictors included age, assigned sex, laboratories, medications and infusions, input/output, Glasgow Coma Scale (GCS) components, and time-series vitals (heart rate, ICP, mean arterial pressure, respiratory rate, and temperature). Each patient ICU stay was divided into successive 95-minute timeblocks. For each timeblock, the model was trained on nontime-varying covariates as well as on 12 observations of time-varying covariates at 5-minute intervals and asked to predict the 5-minute median ICP 30 minutes after the last observed ICP value. Data from 931 patients with ICP monitoring in the eICU dataset were extracted (46,207 timeblocks). The root mean squared error was 4.51 mm Hg in the eICU test set and 3.56 mm Hg in the MIMIC-III dataset. The most important variables driving ICP prediction were previous ICP history, patients' temperature, weight, serum creatinine, age, GCS, and hemodynamic parameters. IntraCranial pressure prediction AlgoRithm using machinE learning, an ensemble machine learning model, trained to predict the ICP of a patient 30 minutes in the future based on baseline characteristics and vitals data from the past hour showed promising predictive performance including in an external validation dataset.

Investigation of statistical methods used in prognostic prediction models for obstetric care: A 10 year-span cross-sectional study.

Obstetric care is a highly active area in the development and application of prognostic prediction models. The development and validation of these models often require the utilization of advanced statistical techniques. However, failure to adhere to rigorous methodological standards could greatly undermine the reliability and trustworthiness of the resultant models. Consequently, the aim of our study was to examine the current statistical practices employed in obstetric care and offer recommendations to enhance the utilization of statistical methods in the development of prognostic prediction models. We conducted a cross-sectional survey using a sample of studies developing or validating prognostic prediction models for obstetric care published in a 10-year span (2011-2020). A structured questionnaire was developed to investigate the statistical issues in five domains, including model derivation (predictor selection and algorithm development), model validation (internal and external), model performance, model presentation, and risk threshold setting. On the ground of survey results and existing guidelines, a list of recommendations for statistical methods in prognostic models was developed. A total of 112 eligible studies were included, with 107 reporting model development and five exclusively reporting external validation. During model development, 58.9% of the studies did not include any form of validation. Of these, 46.4% used stepwise regression in a crude manner for predictor selection, while two-thirds made decisions on retaining or dropping candidate predictors solely based on p-values. Additionally, 26.2% transformed continuous predictors into categorical variables, and 80.4% did not consider nonlinear relationships between predictors and outcomes. Surprisingly, 94.4% of the studies did not examine the correlation between predictors. Moreover, 47.1% of the studies did not compare population characteristics between the development and external validation datasets, and only one-fifth evaluated both discrimination and calibration. Furthermore, 53.6% of the studies did not clearly present the model, and less than half established a risk threshold to define risk categories. In light of these findings, 10 recommendations were formulated to promote the appropriate use of statistical methods. The use of statistical methods is not yet optimal. Ten recommendations were offered to assist the statistical methods of prognostic prediction models in obstetric care.

Machine learning models to predict the warfarin discharge dosage using clinical information of inpatients from South Korea

As warfarin has a narrow therapeutic window and obvious response variability among individuals, it is difficult to rapidly determine personalized warfarin dosage. Adverse drug events(ADE) resulting from warfarin overdose can be critical, so that typically physicians adjust the warfarin dosage through the INR monitoring twice a week when starting warfarin. Our study aimed to develop machine learning (ML) models that predicts the discharge dosage of warfarin as the initial warfarin dosage using clinical data derived from electronic medical records within 2 days of hospitalization. During this retrospective study, adult patients who were prescribed warfarin at Asan Medical Center (AMC) between January 1, 2018, and October 31, 2020, were recruited as a model development cohort (n = 3168). Additionally, we created an external validation dataset (n = 891) from a Medical Information Mart for Intensive Care III (MIMIC-III). Variables for a model prediction were selected based on the clinical rationale that turned out to be associated with warfarin dosage, such as bleeding. The discharge dosage of warfarin was used the study outcome, because we assumed that patients achieved target INR at discharge. In this study, four ML models that predicted the warfarin discharge dosage were developed. We evaluated the model performance using the mean absolute error (MAE) and prediction accuracy. Finally, we compared the accuracy of the predictions of our models and the predictions of physicians for 40 data point to verify a clinical relevance of the models. The MAEs obtained using the internal validation set were as follows: XGBoost, 0.9; artificial neural network, 0.9; random forest, 1.0; linear regression, 1.0; and physicians, 1.3. As a result, our models had better prediction accuracy than the physicians, who have difficulty determining the warfarin discharge dosage using clinical information obtained within 2 days of hospitalization. We not only conducted the internal validation but also external validation. In conclusion, our ML model could help physicians predict the warfarin discharge dosage as the initial warfarin dosage from Korean population. However, conducting a successfully external validation in a further work is required for the application of the models.

Diagnosis and detection of pneumonia using weak-label based on X-ray images: a multi-center study

PurposeDevelopment and assessment the deep learning weakly supervised algorithm for the classification and detection pneumonia via X-ray.MethodsThis retrospective study analyzed two publicly available dataset that contain X-ray images of pneumonia cases and normal cases. The first dataset from Guangzhou Women and Children’s Medical Center. It contains a total of 5,856 X-ray images, which are divided into training, validation, and test sets with 8:1:1 ratio for algorithm training and testing. The deep learning algorithm ResNet34 was employed to build diagnostic model. And the second public dataset were collated by researchers from Qatar University and the University of Dhaka along with collaborators from Pakistan and Malaysia and some medical doctors. A total of 1,300 images of COVID-19 positive cases, 1,300 normal images and 1,300 images of viral pneumonia for external validation. Class activation map (CAM) were used to location the pneumonia lesions.ResultsThe ResNet34 model for pneumonia detection achieved an AUC of 0.9949 [0.9910–0.9981] (with an accuracy of 98.29% a sensitivity of 99.29% and a specificity of 95.57%) in the test dataset. And for external validation dataset, the model obtained an AUC of 0.9835[0.9806–0.9864] (with an accuracy of 94.62%, a sensitivity of 92.35% and a specificity of 99.15%). Moreover, the CAM can accurately locate the pneumonia area.ConclusionThe deep learning algorithm can accurately detect pneumonia and locate the pneumonia area based on weak supervision information, which can provide potential value for helping radiologists to improve their accuracy of detection pneumonia patients through X-ray images.

Retinal Photograph-based Deep Learning System for Detection of Thyroid-Associated Ophthalmopathy

Background: The diagnosis of thyroid-associated ophthalmopathy (TAO) usually requires a comprehensive examination, including clinical symptoms, radiological examinations, and blood tests. Therefore, cost-effective and noninvasive methods for the detection of TAO are needed. This study aimed to establish a deep learning-based system to detect TAO based on retinal photographs. Materials and methods: The multicenter observational study included retinal photographs taken from TAO patients and normal participants in 2 hospitals in China. Forty-five-degree retinal photographs, centered on the midpoint between the optic disc and the macula, were captured by trained ophthalmologists. The authors first trained a convolutional neural network model to identify TAO using data collected from one hospital. After internal validation, the model was further evaluated in another hospital as an external validation data set. Results: The study included 1182 retinal photographs of 708 participants for model development, and 365 retinal photographs (189 participants) were obtained as the external validation data set. In the internal validation, the area under the receiver operator curve was 0.900 (95% CI: 0.889–0.910) and the accuracy was 0.860 (95% CI: 0.849–0.869). In the external data set, the model reached an area under the curve of 0.747 (95% CI: 0.728–0.763) and achieved an accuracy of 0.709 (95% CI: 0.690–0.724). Conclusions: Deep learning-based systems may be promising for identifying TAO in normal subjects using retinal fundus photographs. It may serve as a cost-effective and noninvasive method to detect TAO in the future.