Independent Test Group Research Articles

Use of Artificial Intelligence to Predict Outcomes in Mild Aortic Valve Stenosis

Abstract Background and Aims Aortic stenosis (AS) is a common and progressive disease which, if left untreated, results in increased morbidity and mortality. Monitoring and follow-up care can be challenging due to significant variability in disease progression. This study aimed to develop machine learning models to predict the risks of disease progression and mortality in patients with mild aortic stenosis. Methods A comprehensive database including 9611 patients with serial transthoracic echocardiograms was collected from a single institution across 3 clinical sites. The dataset included parameters from echocardiograms, electrocardiograms, laboratory values, and diagnosis codes. Data from a single clinical site was preserved as an independent test group. Machine learning models were trained to identify progression to severe stenosis and all-cause mortality and tested in their performance for endpoints at 2 and 5 years. Results In the independent test group, the AS progression model differentiated those with progression to severe AS within 2 and 5 years with an area under the curve (AUC) of 0.86 for both. The feature of greatest importance was aortic valve mean gradient, followed by other valve hemodynamic measurements including valve area and dimensionless index. The mortality model identified those with mortality within 2 years and 5 years with an AUC of 0.84 and 0.87, respectively. Smaller reduced-input validation models had similarly robust findings. Conclusions Machine learning models can be used in patients with mild AS to identify those at high risk for disease progression and mortality. Implementation of such models may facilitate real-time, patient-specific follow-up recommendations.

Circulating metabolites associate with the QT interval in individuals without prevalent cardiovascular disease

Abstract Background Patients with metabolic syndrome have a higher prevalence of heart-rate corrected QT (QTc) interval prolongation. Despite this the effect of circulating metabolites on the QTc is largely unknown and limited to small studies. Purpose To test for association of circulating metabolites with the QTc interval in the UK Biobank study. Methods Participants were identified from the UK Biobank with same day plasma metabolite profiling and electrocardiograms. All 249 metabolites measured underwent quality control and log transformation to account for rightward skew and zero values. 149 metabolites with inter-correlations >0.7 were excluded. Participants taking QTc-prolonging medication or with a history of ischaemic heart disease or heart failure up to 6 months after the visit, were excluded. Participants were allocated to a training group (N=21,610) or an independent test group (N=5,304), if samples were obtained at recruitment or at a second visit, respectively. In the training group, each metabolite was tested separately for association with the QTc in linear regression analyses, adjusted for clinical covariates (age, sex, fasting time, body mass index (BMI), systolic blood pressure (SBP), smoking status, diabetes, lipid-lowering drug use and dietary factors). Significant metabolites (P<5x10-4) underwent validation in the test group. Sex-stratified analyses were also performed. In the training group, variables in two models underwent regularisation by LASSO regression (10-fold cross-validation) to reduce multicollinearity and downweigh less important variables; 1) clinical variables, 2) clinical variables and validated metabolites. Coefficients were used for QTc prediction in the test group. Significance between the models was evaluated by fisher’s transformation of R-squared statistics. Results In the per metabolite multivariate analysis, 56 metabolites were associated with the QTc interval. 19 metabolites validated in the test group. For 14 of these, absolute effect sizes were >5ms when comparing individuals in the top decile of the metabolite distribution verses the bottom (Figure 1), including the ketone body 3-hydroxybutyrate (9ms), and omega-6 fatty acid to total fatty acid ratio (7.2ms). For associations in males and females separately, significant effect size differences (P<0.05) were observed for 9 metabolites (Figure 2). In the training group, the combined LASSO model selected 17 (of 19) metabolites along with age, sex, SBP and BMI. In the test group, this model R-squared was 0.115 compared with 0.081 for clinical covariates alone, representing a significant 41.9% increase in QTc variation explained (P=0.002). Conclusions This study demonstrates clinically relevant metabolite associations with the QTc in individuals without cardiovascular disease and explain a significant proportion of QTc variation. Further investigation is warranted to test for direct effects on cardiac electrophysiology and proarrhythmic potential.Validated metabolitesSignificant sex differences

The Effects of Adverse Life Events on Brain Development in the ABCD Study®: A Propensity-weighted Analysis.

Longitudinal studies of the effects of adversity on human brain development are complicated by the association of stressful events with confounding variables. To counter this bias, we apply a propensity-weighted analysis of the first two years of The Adolescent Brain Cognitive DevelopmentSM (ABCD) Study® data, employing a machine learning analysis weighted by individuals' propensity to experience adversity. Data included 338 resting-state functional connections from 7190 youth (46% female), divided into a training group (80%) and an independent testing group (20%). Propensity scores were computed using 390 variables to balance across two-year adverse life event exposures. Using elastic net regularization with and without inverse propensity weighting, we developed linear models in which changes in functional connectivity of brain connections during the two-year period served as predictors of the number of adverse events experienced during that same period. Haufe's method was applied to forward-transform the backward prediction models. We also tested whether brain changes associated with adverse events correlated with concomitant changes in internalizing or externalizing behaviors or to academic achievement. In the propensity-weighted analysis, brain development significantly predicted the number of adverse events experienced during that period in both the training group (ρ=0.14, p<0.001) and the independent testing group (ρ=0.10, p<0.001). The predictor indicated a general pattern of decreased functional connectivity between large-scale networks and subcortical brain regions, particularly for cingulo-opercular and sensorimotor networks. These network-to-subcortical functional connectivity decreases inversely associated with the development of internalizing symptoms, suggesting adverse events promoted adaptive brain changes that may buffer against stress-related psychopathology. However, these same functional connections were also associated with poorer grades at the two-year follow-up. Although cortical-subcortical brain developmental responses to adversity potentially shield against stress-induced mood and anxiety disorders, they may be detrimental to other domains such as academic success.

Identification of necroptosis-associated miRNA signature for predicting prognosis and immune landscape in stomach adenocarcinoma

PurposeThis study aims to identify the potential necroptosis related genes (NRGs)-associated miRNAs signature and explore the impact on the prognosis of stomach adenocarcinoma (STAD). MethodsEmploying rigorous methodologies, we utilized univariate Cox, Lasso and multivariate Cox regression analyses to develop a prognostic signature. Kaplan-Meier (K-M) and ROC curves were applied to assess the prognostic value of signature in a training group and an independent test group. Furthermore, we conducted Gene Set Enrichment Analysis (GSEA) for enrichment of tumor-related pathways. The risk score was calculated for each patient based on the expression of miRNAs which were enrolled in the signature. Patients were stratified into high- and low-risk groups. The immune cell infiltration and immunotherapy were compared between the two groups. Finally, the diagnostic potential of the miRNA was explored by RT-qPCR. ResultsWe constructed a prognostic model based on 6 NRGs-associated miRNAs. K-M plots underscored superior survival outcomes in the low-risk group. GSEA results revealed the enrichment of several tumor-related pathways in the high-risk group. Notably, CD8+ T cells, Tregs and activated memory CD4+ T cells exhibited negative correlations with the risk score. Additionally, a few immune checkpoint genes, such as CTLA4, PD1 and PD-L1, were significantly upregulated in the low-risk group. Furthermore, the serum expression levels of all these 6 miRNAs were significantly elevated in STAD patients. ConclusionsOur study identified a robust risk score derived from a signature of 6 NRGs-associated miRNAs, demonstrating high efficacy for prognosis of STAD. These results not only contributed to our understanding of STAD pathogenesis, but also held promise for potential clinical applications, particularly in the realm of personalized immunotherapy for STAD patients.

Automatic Segmentation of Abdominal Aortic Aneurysms from Time-Resolved 3D Ultrasound Images Using Deep Learning.

Abdominal aortic aneurysms (AAAs) are rupture-prone dilatations of the aorta. In current clinical practice, the maximal diameter of AAAs is monitored with 2D ultrasound to estimate their rupture risk. Recent studies have shown that 3-dimensional and mechanical AAA parameters might be better predictors for aneurysm growth and rupture than the diameter. These parameters can be obtained with time-resolved 3D ultrasound (3D+t US), which requires robust and automatic segmentation of AAAs from 3D+t US. This study proposes and validates a deep learning (DL) approach for automatic segmentation of AAAs. 500 AAA patients were included for follow-up 3D+t US imaging, resulting in 2495 3D+t US images. Segmentation masks for model training were obtained using a conventional automatic segmentation algorithm ('nonDL'). Four different DL models were trained and validated by (1) comparison to CT and (2) reader scoring. Performance of the nonDL and different DL segmentation strategies were evaluated by comparing Hausdorff distance, Dice scores, accuracy, sensitivity, and specificity with a sign test. All DL models had higher median Dice scores, accuracy, and sensitivity (all p < 0.003) compared to nonDL segmentation. The full image-resolution model without data augmentation showed the highest median Dice score and sensitivity (p < 0.001). Applying the DL model on an independent test group produced fewer poor segmentation scores of 1 to 2 on a five-point scale (8% for DL, 18% for nonDL). This demonstrates that a robust and automatic segmentation algorithm for segmenting abdominal aortic aneurysms from 3D+t US images was developed, showing improved performance compared to conventional segmentation.

Machine learning models based on multi-parameter MRI radiomics for prediction of molecular glioblastoma: a new study based on the 2021 World Health Organization classification.

The 2021 World Health Organization (WHO) classification considers a histological low grade glioma with specific molecular characteristics as molecular glioblastoma (mGBM). Accurate identification of mGBM will aid in risk stratification of glioma patients. To explore the value of machine learning models based on magnetic resonance imaging (MRI) radiomics features in predicting mGBM. In total, 166 patients histologically diagnosed as low-grade diffuse glioma (WHO II and III) were included in the study. Fifty-three cases were reclassified as mGBM based on molecular status. Four dimensionality reduction methods including distance correlation (DC), gradient boosted decision tree (GBDT), least absolute shrinkage and selection operator (LASSO) and minimal redundancy maximal relevance (MRMR) were used to select the optimal signatures. Six machine learning algorithms including support vector machine (SVM), linear discriminant analysis (LDA), neural network (NN), logistic regression (LR), K-nearest neighbour (KNN) and decision tree (DT) were used to develop the classifiers. The relative SD was used to evaluate the stability of the models, and the area under the curve values in the independent test group were used to evaluate their performances. NN_DC was determined as the optimal classifier due to the highest area under the curve of 0.891 in the test group. The classification accuracy, sensitivity, specificity, positive predictive value and negative predictive value of NN_DC were 0.915, 0.842, 0.950, 0.889 and 0.927, respectively. Machine learning models can predict mGBM non-invasively, which may help to develop personalized treatment strategies for neurosurgeons and provide an effective tool for accurate stratification in clinical trials.

Development of a nomogram based on radiomics and semantic features for predicting chromosome 7 gain/chromosome 10 loss in IDH wild-type histologically low-grade gliomas.

To predict chromosome 7 gain and chromosome 10 loss (+7/-10) in IDH wild-type (IDH-wt) histologically low-grade gliomas (LGG) by machine learning models based on MRI radiomics and semantic features. A total of 122 patients diagnosed as IDH-wt histologically LGG were retrospectively included in this study. The patients were randomly divided into a training group and a test group in a ratio of 7:3. The radiomics features were extracted from axial T1WI, T2WI, FLAIR and CET1 sequences, respectively. The distance correlation (DC) and least absolute shrinkage and selection operator (LASSO) were used to select the radiomics signatures. Three machine learning algorithms including neural network (NN), support vector machine (SVM), and linear discriminant analysis (LDA) were used to construct radiomics models. In addition, a nomogram was developed by combining the optimal radiomics signature with clinical risk factors, and the potential clinical utility of the nomogram was evaluated using decision curve analysis. The LDA+DC model was identified as the optimal classifier among the six radiomics models. Necrosis was determined as a risk factor for +7/-10 in IDH-wt histologically LGG. The nomogram achieved the best performance, with an AUC of 0.854 and an accuracy of 0.778 in the independent test group. The decision curve of the nomogram confirmed its clinical usefulness in a wide range of thresholds. The nomogram combining radiomics and semantic features can predict the +7/-10 status effectively, which may contribute to the risk stratification and individualized treatment planning of patients with IDH-wt histologically LGG.

Clinical evaluation of deep learning-based automatic clinical target volume segmentation: a single-institution multi-site tumor experience.

The large variability in tumor appearance and shape makes manual delineation of the clinical target volume (CTV) time-consuming, and the results depend on the oncologists' experience. Whereas deep learning techniques have allowed oncologists to automate the CTV delineation, multi-site tumor analysis is often lacking in the literature. This study aimed to evaluate the deep learning models that automatically contour CTVs of tumors at various sites on computed tomography (CT) images from objective and subjective perspectives. 577 patients were selected for the present study, including nasopharyngeal (n = 34), esophageal (n = 40), breast-conserving surgery (BCS) (left-sided, n = 71; right-sided, n = 71), breast-radical mastectomy (BRM) (left-sided, n = 43; right-sided, n = 37), cervical (radical radiotherapy, n = 45; postoperative, n = 85), prostate (n = 42), and rectal (n = 109) carcinomas. Manually delineated CTV contours by radiation oncologists are served as ground truth. Four models were evaluated: Flexnet, Unet, Vnet, and Segresnet, which are commercially available in the medical product "AccuLearning AI model training platform". The data were divided into the training, validation, and testing set at a ratio of 5:1:4. The geometric metrics, including Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD), were calculated for objective evaluation. For subjective assessment, oncologists rated the segmentation contours of the testing set visually. High correlations were observed between automatic and manual contours. Based on the results of the independent test group, most of the patients achieved satisfactory quantitative results (DSC > 0.8), except for patients with esophageal carcinoma (DSC: 0.62-0.64). The subjective review indicated that 82.65% of predicted CTVs scored either as clinically accepting (8.68%) or requiring minor revision (73.97%), and no patients were scored as rejected. This experimental work demonstrated that auto-generated contours could serve as an initial template to help oncologists save time in CTV delineation. The deep learning-based auto-segmentations achieve acceptable accuracy and show the potential to improve clinical efficiency for radiotherapy of a variety of cancer.

Evolutionary learning of selection hyper-heuristics for text classification

This paper introduces an evolutionary model in the scope of automated machine learning. This model is in charge of learning hyper-heuristics that represent selection rules of the form if-then, such that given a dataset for a text classification problem, the hyper-heuristics select the best classification method to use with it, based on the data distribution of the dataset. The evolutionary model starts by building a set of hyper-heuristics using a series of meta-features extracted from a training group of datasets that represent their data distribution. Hyper-heuristics are then evolved using adapted crossover and mutation operators. During the evolution, each hyper-heuristic is evaluated on its performance to classify each dataset in the training group. When the evolutionary process is done, the best hyper-heuristic is selected and evaluated for its generality with an independent test group of datasets. The results show that the best learned hyper-heuristic obtains an average classification performance close to the general optimum, and has a similar performance to the two most popular state-of-the-art automated machine learning systems, but with less computational cost. The approach used by the present model is relevant for automated machine learning in three aspects, the generality of the hyper-heuristics so they could be applied to groups of datasets; the interpretability of the representations that facilitate the understanding of the method selection by non-expert users; and the reduction of computational time and resources to reach a decision. Furthermore, the model extends the applicability of evolutionary computation methods, with their problem-independent properties and their ability to explore search spaces, to tackle new complex problems, such as the decision of the best classifier for a text classification dataset.

A nomogram for predicting the risk of residual stone fragments after ureteroscopy.

The residual stone fragment is a tremendous issue after ureteroscopic lithotripsy and requires urologists to evaluate the condition of patients comprehensively. Our study aimed to construct a nomogram to make a personalized prediction of postoperative residual stone rate (RSR). We implemented a retrospective cohort study in the Department of Urology, Shanghai General Hospital. A total of 277 patients undergoing ureteroscopy (URS) were enrolled in our study. Among them, 186 patients were included in the training group and the remaining 91 patients comprised the testing group. We utilized stepwise forward algorithm and logistic regression analysis to build predictive models and selected the best model based on Akaike's information criterion (AIC). The model was assessed by receiver operating characteristic (ROC) curves and the Hosmer-Lemeshow (HL) test. We also conducted decision curve analysis (DCA) to demonstrate the net benefit of the model. The independent testing group was used to validate the practicability of the nomogram. The severity of hydronephrosis, stone location, the transverse diameter of stone, hypertension, and white blood cell (WBC) were found to be significant predictive variables for RSR after URS. The area under the curve (AUC) of the training group was 0.7203 and that of the testing group was 0.7280. Besides, the nomogram also presented great calibration and accepted net benefit in a wide range of probabilities. Our study achieved a predictive nomogram with excellent application value for urologists to assess RSR and make personalized treatment decisions.

Alzheimer's Precision Neurology: Epigenetics of Cytochrome P450 Genes in Circulating Cell-Free DNA for Disease Prediction and Mechanism.

Precision neurology combines high-throughput technologies and statistical modeling to identify novel disease pathways and predictive biomarkers in Alzheimer's disease (AD). Brain cytochrome P450 (CYP) genes are major regulators of cholesterol, sex hormone, and xenobiotic metabolism, and they could play important roles in neurodegenerative disorders. Increasing evidence suggests that epigenetic factors contribute to AD development. We evaluated cytosine ('CpG')-based DNA methylation changes in AD using circulating cell-free DNA (cfDNA), to which neuronal cells are known to contribute. We investigated CYP-based mechanisms for AD pathogenesis and epigenetic biomarkers for disease detection. We performed a case-control study using 25 patients with AD and 23 cognitively healthy controls using the cfDNA of CYP genes. We performed a logistic regression analysis using the MetaboAnalyst software computer program and a molecular pathway analysis based on epigenetically altered CYP genes using the Cytoscape program. We identified 130 significantly (false discovery rate correction q-value < 0.05) differentially methylated CpG sites within the CYP genes. The top two differentially methylated genes identified were CYP51A1 and CYP2S1. The significant molecular pathways that were perturbed in AD cfDNA were (i) androgen and estrogen biosynthesis and metabolism, (ii) C21 steroid hormone biosynthesis and metabolism, and (iii) arachidonic acid metabolism. Existing evidence suggests a potential role of each of these biochemical pathways in AD pathogenesis. Next, we randomly divided the study group into discovery and validation sub-sets, each consisting of patients with AD and control patients. Regression models for AD prediction based on CYP CpG methylation markers were developed in the discovery or training group and tested in the independent validation group. The CYP biomarkers achieved a high predictive accuracy. After a 10-fold cross-validation, the combination of cg17852385/cg23101118 + cg14355428/cg22536554 achieved an AUC (95% CI) of 0.928 (0.787~1.00), with 100% sensitivity and 92.3% specificity for AD detection in the discovery group. The performance remained high in the independent validation or test group, achieving an AUC (95% CI) of 0.942 (0.905~0.979) with a 90% sensitivity and specificity. Our findings suggest that the epigenetic modification of CYP genes may play an important role in AD pathogenesis and that circulating CYP-based cfDNA biomarkers have the potential to accurately and non-invasively detect AD.

Prediction of pathologic complete response to neoadjuvant systemic therapy in triple negative breast cancer using deep learning on multiparametric MRI

Triple-negative breast cancer (TNBC) is an aggressive subtype of breast cancer. Neoadjuvant systemic therapy (NAST) followed by surgery are currently standard of care for TNBC with 50-60% of patients achieving pathologic complete response (pCR). We investigated ability of deep learning (DL) on dynamic contrast enhanced (DCE) MRI and diffusion weighted imaging acquired early during NAST to predict TNBC patients’ pCR status in the breast. During the development phase using the images of 130 TNBC patients, the DL model achieved areas under the receiver operating characteristic curves (AUCs) of 0.97 ± 0.04 and 0.82 ± 0.10 for the training and the validation, respectively. The model achieved an AUC of 0.86 ± 0.03 when evaluated in the independent testing group of 32 patients. In an additional prospective blinded testing group of 48 patients, the model achieved an AUC of 0.83 ± 0.02. These results demonstrated that DL based on multiparametric MRI can potentially differentiate TNBC patients with pCR or non-pCR in the breast early during NAST.

Preoperative prediction and histological stratification of intracranial solitary fibrous tumours by machine-learning models

To explore the effectiveness and feasibility of machine-learning models based on magnetic resonance imaging (MRI) radiomics features in differentiating intracranial solitary fibrous tumour (ISFT) from angiomatous meningioma (AM) and stratifying ISFT histologically. This study retrospectively recruited 268 patients with a histological diagnosis of ISFT (n=120) or AM (n=148), and 116 of the ISFT patients were used for stratified analysis of histological grade. The radiomics features were extracted from axial T1-weighted imaging (WI), T2WI and contrast-enhanced T1WI sequences. All patients were assigned randomly to the training group and test group in a ratio of 7:3. The models were optimised by 10-fold cross-validation in the training group, and the independent test group was used for further testing of the models. The performances of machine-learning models based on radiomics, clinical, and fusion features in predicting and stratifying ISFT were evaluated. ISFT and AM differed significantly in terms of age, tumour shape, enhancement pattern, and margin. There was no significant difference in the clinical characteristics between World Health Organization (WHO) grade II and WHO grade III ISFT. When used to differentiate ISFT from AM, the area under the curve (AUC) values of the machine-learning models based on radiomics, clinical, and fusion features in the test group were 0.917, 0.923 and 0.950, respectively. When used for histological stratification of ISFT, the model based on the radiomics signature achieved an AUC value of 0.786 in the test group. Machine-learning models can contribute in the prediction and histological stratification of ISFT non-invasively, which can help clinical differential diagnosis and treatment decisions.

Obstructive Sleep Apnea Detection Based on Sleep Sounds via Deep Learning.

This study aimed to propose a novel deep-learning method for automatic sleep apneic event detection and thus to estimate the apnea hypopnea index (AHI) and identify obstructive sleep apnea (OSA) in an event-by-event manner solely based on sleep sounds obtained by a noncontact audio recorder. We conducted a cross-sectional study of participants with habitual snoring or heavy breathing sounds during sleep to train and test a deep convolutional neural network named OSAnet for the detection of OSA based on sleep sounds. Polysomnography (PSG) was conducted, and sleep sounds were recorded simultaneously in a regular room without noise attenuation. The study was conducted in two phases. In phase one, eligible participants were enrolled and randomly allocated into training and validation groups for deep learning algorithm development. In phase two, eligible patients were enrolled in a test group for algorithm assessment. Sensitivity, specificity, accuracy, unweighted Cohen kappa coefficient (κ) and the area under the curve (AUC) were calculated using PSG as the reference standard. A total of 135 participants were randomly divided into a training group (n, 116) and a validation group (n, 19). An independent test group of 59 participants was subsequently enrolled. Our algorithm achieved a precision of 0.81 and sensitivity of 0.78 in the test group for overall sleep event detection. The algorithm exhibited robust diagnostic performance to identify severe cases with a sensitivity of 95.6% and specificity of 91.6%. Our results showed that a deep learning algorithm based on sleep sounds recorded by a noncontact voice recorder served as a feasible tool for apneic event detection and OSA identification. This technique may hold promise for OSA assessment in the community in a relatively comfortable and low-cost manner. Further studies to develop a tool based on a home-based setting are warranted.

One 3D VOI-based deep learning radiomics strategy, clinical model and radiologists for predicting lymph node metastases in pancreatic ductal adenocarcinoma based on multiphasic contrast-enhanced computer tomography.

PurposeWe designed to construct one 3D VOI-based deep learning radiomics strategy for identifying lymph node metastases (LNM) in pancreatic ductal adenocarcinoma on the basis of multiphasic contrast-enhanced computer tomography and to assist clinical decision-making.MethodsThis retrospective research enrolled 139 PDAC patients undergoing pre-operative arterial phase and venous phase scanning examination between 2015 and 2021. A primary group (training group and validation group) and an independent test group were divided. The DLR strategy included three sections. (1) Residual network three dimensional-18 (Resnet 3D-18) architecture was constructed for deep learning feature extraction. (2) Least absolute shrinkage and selection operator model was used for feature selection. (3) Fully connected network served as the classifier. The DLR strategy was applied for constructing different 3D CNN models using 5-fold cross-validation. Radiomics scores (Rad score) were calculated for distinguishing the statistical difference between negative and positive lymph nodes. A clinical model was constructed by combining significantly different clinical variables using univariate and multivariable logistic regression. The manifestation of two radiologists was detected for comparing with computer-developed models. Receiver operating characteristic curves, the area under the curve, accuracy, precision, recall, and F1 score were used for evaluating model performance.ResultsA total of 45, 49, and 59 deep learning features were selected via LASSO model. No matter in which 3D CNN model, Rad score demonstrated the deep learning features were significantly different between non-LNM and LNM groups. The AP+VP DLR model yielded the best performance in predicting status of lymph node in PDAC with an AUC of 0.995 (95% CI:0.989-1.000) in training group; an AUC of 0.940 (95% CI:0.910-0.971) in validation group; and an AUC of 0.949 (95% CI:0.914-0.984) in test group. The clinical model enrolled the histological grade, CA19-9 level and CT-reported tumor size. The AP+VP DLR model outperformed AP DLR model, VP DLR model, clinical model, and two radiologists.ConclusionsThe AP+VP DLR model based on Resnet 3D-18 demonstrated excellent ability for identifying LNM in PDAC, which could act as a non-invasive and accurate guide for clinical therapeutic strategies. This 3D CNN model combined with 3D tumor segmentation technology is labor-saving, promising, and effective.

Development of a Machine Learning Model to Discriminate Mild Cognitive Impairment Subjects from Normal Controls in Community Screening.

Background: Mild cognitive impairment (MCI) is a transitional stage between normal aging and probable Alzheimer’s disease. It is of great value to screen for MCI in the community. A novel machine learning (ML) model is composed of electroencephalography (EEG), eye tracking (ET), and neuropsychological assessments. This study has been proposed to identify MCI subjects from normal controls (NC). Methods: Two cohorts were used in this study. Cohort 1 as the training and validation group, includes184 MCI patients and 152 NC subjects. Cohort 2 as an independent test group, includes 44 MCI and 48 NC individuals. EEG, ET, Neuropsychological Tests Battery (NTB), and clinical variables with age, gender, educational level, MoCA-B, and ACE-R were selected for all subjects. Receiver operating characteristic (ROC) curves were adopted to evaluate the capabilities of this tool to classify MCI from NC. The clinical model, the EEG and ET model, and the neuropsychological model were compared. Results: We found that the classification accuracy of the proposed model achieved 84.5 ± 4.43% and 88.8 ± 3.59% in Cohort 1 and Cohort 2, respectively. The area under curve (AUC) of the proposed tool achieved 0.941 (0.893–0.982) in Cohort 1 and 0.966 (0.921–0.988) in Cohort 2, respectively. Conclusions: The proposed model incorporation of EEG, ET, and neuropsychological assessments yielded excellent classification performances, suggesting its potential for future application in cognitive decline prediction.

Prediction of subsequent osteoporotic vertebral compression fracture on CT radiography via deep learning

AbstractCombination of computed tomography (CT) radiography and deep learning to predict subsequent osteoporotic vertebral compression fracture (OVCF) has not been reported. To do so, we analyzed retrospectively CT images from 103 patients who experienced twice OVCF in Tongji Hospital from 2011 to 2022. Meanwhile, CT images from 70 age‐matched osteoporotic patients without vertebral fracture were used as the negative control. Convolutional neural network was used for classification and the Adam optimizer combining the momentum and exponentially weighted moving average gradients methods were used to update the weights of the networks. In the prediction model, we split 80% data of each type of the patient as the training group, while the other 20% was held as the independent testing group. We found that the number of subsequent fracture in women is higher than that in men (81 vs. 22). Additionally, the incidence rate of adjacent vertebral fracture is higher than that of remote vertebral fracture (64.1 vs. 35.9%), while the onset time of the former was 11.9 ± 12.8 months, significantly less than 22.3 ± 18.2 months of the latter (p &lt; .001). For the prediction of subsequent fracture, our model attained .839 of accuracy and .883 of receiver operating characteristic–area under curve on the whole testing dataset. Furthermore, our model gained .867 and .719 of accuracy on the single‐class testing dataset separated from the former, .817 of accuracy on the independent test. In conclusion, we managed to generate a deep learning‐based model, which is able to predict subsequent OVCF in a precise and unbiased way just using CT images.

Artificial Intelligence and Circulating Cell-Free DNA Methylation Profiling: Mechanism and Detection of Alzheimer's Disease.

Background: Despite extensive efforts, significant gaps remain in our understanding of Alzheimer’s disease (AD) pathophysiology. Novel approaches using circulating cell-free DNA (cfDNA) have the potential to revolutionize our understanding of neurodegenerative disorders. Methods: We performed DNA methylation profiling of cfDNA from AD patients and compared them to cognitively normal controls. Six Artificial Intelligence (AI) platforms were utilized for the diagnosis of AD while enrichment analysis was used to elucidate the pathogenesis of AD. Results: A total of 3684 CpGs were significantly (adj. p-value < 0.05) differentially methylated in AD versus controls. All six AI algorithms achieved high predictive accuracy (AUC = 0.949–0.998) in an independent test group. As an example, Deep Learning (DL) achieved an AUC (95% CI) = 0.99 (0.95–1.0), with 94.5% sensitivity and specificity. Conclusion: We describe numerous epigenetically altered genes which were previously reported to be differentially expressed in the brain of AD sufferers. Genes identified by AI to be the best predictors of AD were either known to be expressed in the brain or have been previously linked to AD. We highlight enrichment in the Calcium signaling pathway, Glutamatergic synapse, Hedgehog signaling pathway, Axon guidance and Olfactory transduction in AD sufferers. To the best of our knowledge, this is the first reported genome-wide DNA methylation study using cfDNA to detect AD.

Using Deep Learning Radiomics to Distinguish Cognitively Normal Adults at Risk of Alzheimer's Disease From Normal Control: An Exploratory Study Based on Structural MRI.

ObjectivesWe proposed a novel deep learning radiomics (DLR) method to distinguish cognitively normal adults at risk of Alzheimer’s disease (AD) from normal control based on T1-weighted structural MRI images.MethodsIn this study, we selected MRI data from the Alzheimer’s Disease Neuroimaging Initiative Database (ADNI), which included 417 cognitively normal adults. These subjects were divided into 181 individuals at risk of Alzheimer’s disease (preAD group) and 236 normal control individuals (NC group) according to standard uptake ratio >1.18 calculated by amyloid Positron Emission Tomography (PET). We further divided the preaAD group into APOE+ and APOE− subgroups according to whether APOE ε4 was positive or not. All data sets were divided into one training/validation group and one independent test group. The proposed DLR method included three steps: (1) the pre-training of basic deep learning (DL) models, (2) the extraction, selection and fusion of DLR features, and (3) classification. The support vector machine (SVM) was used as the classifier. In the comparative experiments, we compared our proposed DLR method with three existing models: hippocampal model, clinical model, and traditional radiomics model. Ten-fold cross-validation was performed with 100 time repetitions.ResultsThe DLR method achieved the best classification performance between preAD and NC than other models with an accuracy of 89.85% ± 1.12%. In comparison, the accuracies of the other three models were 72.44% ± 1.37%, 82.00% ± 4.09% and 79.65% ± 2.21%. In addition, the DLR model also showed the best classification performance (85.45% ± 9.04% and 92.80% ± 2.61%) in the subgroup experiment.ConclusionThe results showed that the DLR method provided a potentially clinical value to distinguish preAD from NC.

Abstract P1-08-03: Deep learning for early prediction of neoadjuvant chemotherapy response in triple negative breast cancers

Abstract Introduction: Neoadjuvant chemotherapy (NACT) is becoming standard of care for presurgical treatment of triple negative breast cancer (TNBC) patients. Achievement of pathological complete response (pCR) after NACT is associated with improved outcomes. There is currently an unmet need in development of imaging and clinical tools for prediction of pCR to NACT in TNBC. We investigated use of deep learning convolution neural networks (CNNs) for early prediction of pCR in a TNBC cohort on the basis of MRI acquired before the initiation and at the midpoint, after completion of four cycles of NACT (C4). Materials and Methods: Baseline and C4 MRIs of 112 TNBC patients were collected from an ongoing prospective clinical trial (NCT02276443). Four patients were excluded because they underwent different treatment for the second regimen. Among the 108 patients, 52 patients (48%) had pCR confirmed at surgery. Positive enhancement integral (PEI) derived from the early phases of DCE MRI, and apparent diffusion coefficients (ADC) derived from DWI MRI (b = 100 and 800 s/mm2), were used for our investigation. The images were aligned and the tumor regions were cropped from all images. All tumor patches were normalized between [0, 1], and were padded to form matrices of the same size of 192×192×64 for PEI, or the size of 192×192×16 for ADC. The CNN was constructed using stacked 3D convolution and MaxPooling layers. It consisted of up to four channels for the inputs (baseline and C4 PEI and ADC). Features extracted from each channel were concatenated and regressed for pCR prediction via three densely connected layers. Binary cross-entropy was used as the loss function for CNN training, and the loss was optimized using an Adam optimizer with the initial learning rate of 0.0001. Because of the currently limited sample size, four-fold cross-validation was used for CNN training and evaluation. The patients were divided into four groups, each group had 27 patients and the pCR:non-pCR ratio was controlled as 13:14. For each fold, one group was reserved as the independent testing group, and the other three groups were combined for network training and internal validation. Receiver operating characteristic (ROC) curve was plotted for each fold of testing, and area under the curve (AUC) was calculated. Final performance of the CNN was determined by averaging the AUCs of the four testing folds. Additionally, to test the prediction efficacy of each input, we trained the CNN under the same settings but used PEI or ADC only as input, and the results were compared. Results: The CNN trained with PEI only achieved an average AUC of 0.65 ± 0.09. The second CNN trained with ADC only achieved an average AUC of 0.72 ± 0.07. The third CNN trained with both PEI and ADC achieved an average AUC of 0.73 ± 0.06. Conclusion and Discussion: Using baseline and mid-treatment MRIs, deep learning CNN showed promising performance to predict pCR in the early course of NACT. The prediction AUC for the independent testing groups was largely improved by using ADC to train the network, indicating that ADC can have more critical information than PEI in assisting pCR prediction during the early course of NACT. Future work includes curation of a larger patient data for network training and evaluation to improve the prediction performance and further validate generalization of the network. We will also explore more advanced network structures, through which the prediction performance can be improved. Four-fold cross-validation AUCs of the network using different data as inputs.PEIADCPEI+ADCFold 10.570.640.66Fold 20.760.800.77Fold 30.660.700.68Fold 40.590.740.79Average0.65 ± 0.090.72 ± 0.070.73 ± 0.06 Citation Format: Zijian Zhou, Nabil A Elshafeey, David E Rauch, Beatriz E Adrada, Rosalind P Candelaria, Mary S Guirguis, Wei Yang, Medine Boge, Rania M Mohamed, Gary J Whitman, Deanna L Lane, Huong C Le-Petross, Jessica WT Leung, Lumarie Santiago, Marion E Scoggins, David A Spak, Miral M Patel, Frances Perez, Debu Tripathy, Vicente Valero, Clinton Yam, Stacy Moulder, Jason B White, Jong Bum Son, Mark D Pagel, Jingfei Ma. Deep learning for early prediction of neoadjuvant chemotherapy response in triple negative breast cancers [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr P1-08-03.