Deep Learning Model for Prediction of Bronchopulmonary Dysplasia in Preterm Infants Using Chest Radiographs.

Bronchopulmonary dysplasia (BPD) is common in preterm infants and may result in pulmonary vascular disease, compromising lung function. This study aimed to employ artificial intelligence (AI) techniques to help physicians accurately diagnose BPD in preterm infants in a timely and efficient manner. This retrospective study involves two datasets: a lung region segmentation dataset comprising 1491 chest radiographs of infants, and a BPD prediction dataset comprising 1021 chest radiographs of preterm infants. Transfer learning of a pre-trained machine learning model was employed for lung region segmentation and image fusion for BPD prediction to enhance the performance of the AI model. The lung segmentation model uses transfer learning to achieve a dice score of 0.960 for preterm infants with ≤ 168 h postnatal age. The BPD prediction model exhibited superior diagnostic performance compared to that of experts and demonstrated consistent performance for chest radiographs obtained at ≤ 24 h postnatal age, and those obtained at 25 to 168 h postnatal age. This study is the first to use deep learning on preterm chest radiographs for lung segmentation to develop a BPD prediction model with an early detection time of less than 24 h. Additionally, this study compared the model's performance according to both NICHD and Jensen criteria for BPD. Results demonstrate that the AI model surpasses the diagnostic accuracy of experts in predicting lung development in preterm infants.

A computed tomography radiomics-based model for predicting osteoporosis after breast cancer treatment.

Many treatments against breast cancer decrease the level of estrogen in blood, resulting in bone loss, osteoporosis and fragility fractures in breast cancer patients. This retrospective study aimed to evaluate a novel opportunistic screening for cancer treatment-induced bone loss (CTIBL) in breast cancer patients using CT radiomics. Between 2011 and 2021, a total of 412 female breast cancer patients who received treatment and were followed up in our institution, had post-treatment dual-energy X-ray absorptiometry (DXA) examination of the lumbar vertebrae and had post-treatment chest CT scan that encompassed the L1 vertebra, were included in this study. Results indicated that the T-score of L1 vertebra had a strongly positive correlation with the average T-score of L1-L4 vertebrae derived from DXA (r = 0.91, p < 0.05). On multivariable analysis, four clinical variables (age, body weight, menopause status, aromatase inhibitor exposure duration) and three radiomic features extracted from the region of interest of L1 vertebra (original_firstorder_RootMeanSquared, wavelet.HH_glcm_InverseVariance, and wavelet.LL_glcm_MCC) were selected for building predictive models of L1 T-score and bone health. The predictive model combining clinical and radiomic features showed the greatest adjusted R2 value (0.557), sensitivity (83.6%), specificity (74.2%) and total accuracy (79.4%) compared to models that relied solely on clinical data, radiomic features, or Hounsfield units. In conclusion, the clinical-radiomic predictive model may be used as an opportunistic screening tool for early identification of breast cancer survivors at high risk of CTIBL based on non-contrast CT images of the L1 vertebra, thereby facilitating early intervention for osteoporosis.

Federated machine learning for predicting acute kidney injury in critically ill patients: a multicenter study in Taiwan.

Purpose: To address the contentious data sharing across hospitals, this study adopted a novel approach, federated learning (FL), to establish an aggregate model for acute kidney injury (AKI) prediction in critically ill patients in Taiwan. Methods: This study used data from the Critical Care Database of Taichung Veterans General Hospital (TCVGH) from 2015 to 2020 and electrical medical records of the intensive care units (ICUs) between 2018 and 2020 of four referral centers in different areas across Taiwan. AKI prediction models were trained and validated thereupon. An FL-based prediction model across hospitals was then established. Results: The study included 16,732 ICU admissions from the TCVGH and 38,424 ICU admissions from the other four hospitals. The complete model with 60 features and the parsimonious model with 21 features demonstrated comparable accuracies using extreme gradient boosting, neural network (NN), and random forest, with an area under the receiver-operating characteristic (AUROC) curve of approximately 0.90. The Shapley Additive Explanations plot demonstrated that the selected features were the key clinical components of AKI for critically ill patients. The AUROC curve of the established parsimonious model for external validation at the four hospitals ranged from 0.760 to 0.865. NN-based FL slightly improved the model performance at the four centers. Conclusion: A reliable prediction model for AKI in ICU patients was developed with a lead time of 24 h, and it performed better when the novel FL platform across hospitals was implemented. Supplementary Information: The online version contains supplementary material available at 10.1007/s13755-023-00248-5.

Analytics of self-regulated learning scaffolding: effects on learning processes.

Self-regulated learning (SRL) is the ability to regulate cognitive, metacognitive, motivational, and emotional states while learning and is posited to be a strong predictor of academic success. It is therefore important to provide learners with effective instructions to promote more meaningful and effective SRL processes. One way to implement SRL instructions is through providing real-time SRL scaffolding while learners engage with a task. However, previous studies have tended to focus on fixed scaffolding rather than adaptive scaffolding that is tailored to student actions. Studies that have investigated adaptive scaffolding have not adequately distinguished between the effects of adaptive and fixed scaffolding compared to a control condition. Moreover, previous studies have tended to investigate the effects of scaffolding at the task level rather than shorter time segments-obscuring the impact of individual scaffolds on SRL processes. To address these gaps, we (a) collected trace data about student activities while working on a multi-source writing task and (b) analyzed these data using a cutting-edge learning analytic technique- ordered network analysis (ONA)-to model, visualize, and explain how learners' SRL processes changed in relation to the scaffolds. At the task level, our results suggest that learners who received adaptive scaffolding have significantly different patterns of SRL processes compared to the fixed scaffolding and control conditions. While not significantly different, our results at the task segment level suggest that adaptive scaffolding is associated with earlier engagement in SRL processes. At both the task level and task segment level, those who received adaptive scaffolding, compared to the other conditions, exhibited more task-guided learning processes such as referring to task instructions and rubrics in relation to their reading and writing. This study not only deepens our understanding of the effects of scaffolding at different levels of analysis but also demonstrates the use of a contemporary learning analytic technique for evaluating the effects of different kinds of scaffolding on learners' SRL processes.

Do teaching staff trust stakeholders and tools in learning analytics? A mixed methods study.

Learning analytics (LA) has gained increasing attention for its potential to improve different educational aspects (e.g., students' performance and teaching practice). The existing literature identified some factors that are associated with the adoption of LA in higher education, such as stakeholder engagement and transparency in data use. The broad literature on information systems also emphasizes the importance of trust as a critical predictor of technology adoption. However, the extent to which trust plays a role in the adoption of LA in higher education has not been examined in detail in previous research. To fill this literature gap, we conducted a mixed method (survey and interviews) study aimed to explore how much teaching staff trust LA stakeholders (e.g., higher education institutions or third-parties) and LA technology, as well as the trust factors that could hinder or enable adoption of LA. The findings show that the teaching staff had a high level of trust in the competence of higher education institutions and the usefulness of LA; however, the teaching staff had a low level of trust in third parties that are involved in LA (e.g., external technology vendors) in terms of handling privacy and ethics-related issues. They also had a low level of trust in data accuracy due to issues such as outdated data and lack of data governance. The findings have strategic implications for institutional leaders and third parties in the adoption of LA by providing recommendations to increase trust, such as, improving data accuracy, developing policies for data sharing and ownership, enhancing the consent-seeking process, and establishing data governance guidelines. Therefore, this study contributes to the literature on the adoption of LA in HEIs by integrating trust factors.

A non-genetic engineering platform for rapidly generating and expanding cancer-specific armed T cells.

BACKGROUND: Cancer-specific adoptive T cell therapy has achieved successful milestones in multiple clinical treatments. However, the commercial production of cancer-specific T cells is often hampered by laborious cell culture procedures, the concern of retrovirus-based gene transfection, or insufficient T cell purity. METHODS: In this study, we developed a non-genetic engineering technology for rapidly manufacturing a large amount of cancer-specific T cells by utilizing a unique anti-cancer/anti-CD3 bispecific antibody (BsAb) to directly culture human peripheral blood mononuclear cells (PBMCs). The anti-CD3 moiety of the BsAb bound to the T cell surface and stimulated the differentiation and proliferation of T cells in PBMCs. The anti-cancer moiety of the BsAb provided these BsAb-armed T cells with the cancer-targeting ability, which transformed the naïve T cells into cancer-specific BsAb-armed T cells. RESULTS: With this technology, a large amount of cancer-specific BsAb-armed T cells can be rapidly generated with a purity of over 90% in 7 days. These BsAb-armed T cells efficiently accumulated at the tumor site both in vitro and in vivo. Cytotoxins (perforin and granzyme) and cytokines (TNF-α and IFN-γ) were dramatically released from the BsAb-armed T cells after engaging cancer cells, resulting in a remarkable anti-cancer efficacy. Notably, the BsAb-armed T cells did not cause obvious cytokine release syndrome or tissue toxicity in SCID mice bearing human tumors. CONCLUSIONS: Collectively, the BsAb-armed T cell technology represents a simple, time-saving, and highly safe method to generate highly pure cancer-specific effector T cells, thereby providing an affordable T cell immunotherapy to patients.

Controlling the confounding effect of metabolic gene expression to identify actual metabolite targets in microsatellite instability cancers.

BACKGROUND: The metabolome is the best representation of cancer phenotypes. Gene expression can be considered a confounding covariate affecting metabolite levels. Data integration across metabolomics and genomics to establish the biological relevance of cancer metabolism is challenging. This study aimed to eliminate the confounding effect of metabolic gene expression to reflect actual metabolite levels in microsatellite instability (MSI) cancers. METHODS: In this study, we propose a new strategy using covariate-adjusted tensor classification in high dimensions (CATCH) models to integrate metabolite and metabolic gene expression data to classify MSI and microsatellite stability (MSS) cancers. We used datasets from the Cancer Cell Line Encyclopedia (CCLE) phase II project and treated metabolomic data as tensor predictors and data on gene expression of metabolic enzymes as confounding covariates. RESULTS: The CATCH model performed well, with high accuracy (0.82), sensitivity (0.66), specificity (0.88), precision (0.65), and F1 score (0.65). Seven metabolite features adjusted for metabolic gene expression, namely, 3-phosphoglycerate, 6-phosphogluconate, cholesterol ester, lysophosphatidylethanolamine (LPE), phosphatidylcholine, reduced glutathione, and sarcosine, were found in MSI cancers. Only one metabolite, Hippurate, was present in MSS cancers. The gene expression of phosphofructokinase 1 (PFKP), which is involved in the glycolytic pathway, was related to 3-phosphoglycerate. ALDH4A1 and GPT2 were associated with sarcosine. LPE was associated with the expression of CHPT1, which is involved in lipid metabolism. The glycolysis, nucleotide, glutamate, and lipid metabolic pathways were enriched in MSI cancers. CONCLUSIONS: We propose an effective CATCH model for predicting MSI cancer status. By controlling the confounding effect of metabolic gene expression, we identified cancer metabolic biomarkers and therapeutic targets. In addition, we provided the possible biology and genetics of MSI cancer metabolism.

High-Resolution Genomic Profiling of Liver Cancer Links Etiology With Mutation and Epigenetic Signatures.

BACKGROUND & AIMS: Hepatocellular carcinoma (HCC) is a model of a diverse spectrum of cancers because it is induced by well-known etiologies, mainly hepatitis C virus (HCV) and hepatitis B virus. Here, we aimed to identify HCV-specific mutational signatures and explored the link between the HCV-related regional variation in mutations rates and HCV-induced alterations in genome-wide chromatin organization. METHODS: To identify an HCV-specific mutational signature in HCC, we performed high-resolution targeted sequencing to detect passenger mutations on 64 HCC samples from 3 etiology groups: hepatitis B virus, HCV, or other. To explore the link between the genomic signature and genome-wide chromatin organization we performed chromatin immunoprecipitation sequencing for the transcriptionally permissive H3K4Me3, H3K9Ac, and suppressive H3K9Me3 modifications after HCV infection. RESULTS: Regional variation in mutation rate analysis showed significant etiology-dependent regional mutation rates in 12 genes: LRP2, KRT84, TMEM132B, DOCK2, DMD, INADL, JAK2, DNAH6, MTMR9, ATM, SLX4, and ARSD. We found an enrichment of C->T transversion mutations in the HCV-associated HCC cases. Furthermore, these cases showed regional variation in mutation rates associated with genomic intervals in which HCV infection dictated epigenetic alterations. This signature may be related to the HCV-induced decreased expression of genes encoding key enzymes in the base excision repair pathway. CONCLUSIONS: We identified novel distinct HCV etiology-dependent mutation signatures in HCC associated with HCV-induced alterations in histone modification. This study presents a link between cancer-causing mutagenesis and the increased predisposition to liver cancer in chronic HCV-infected individuals, and unveils novel etiology-specific mechanisms leading to HCC and cancer in general.

Decision tree algorithm predicts hepatocellular carcinoma among chronic hepatitis C patients following viral eradication.

Successful eradication of the hepatitis C virus (HCV) cannot eliminate the risk of hepatocellular carcinoma (HCC). Next-generation RNA sequencing provides comprehensive genomic insights into the pathogenesis of HCC. Artificial intelligence has opened a new era in precision medicine. This study integrated clinical features and genetic biomarkers to establish a machine learning-based HCC model following viral eradication. A prospective cohort of 55 HCV patients with advanced fibrosis, who achieved a sustained virologic response after antiviral therapy, was enrolled. The primary outcome was the occurrence of HCC. The genomic signatures of peripheral blood mononuclear cells (PBMC) were determined by RNA sequencing at baseline and 24 weeks after end-of-treatment. Machine learning algorithms were implemented to extract the predictors of HCC. HCC occurred in 8 of the 55 patients, with an annual incidence of 2.7%. Pretreatment PBMC DEFA1B, HBG2, ADCY4, and posttreatment TAS1R3, ABCA3, and FOSL1 genes were significantly downregulated, while the pretreatment ANGPTL6 gene was significantly upregulated in the HCC group compared to that in the non-HCC group. A gene score derived from the result of the decision tree algorithm can identify HCC with an accuracy of 95.7%. Gene score = TAS1R3 (≥0.63 FPKM, yes/no = 0/1) + FOSL1 (≥0.27 FPKM, yes/no = 0/1) + ABCA3 (≥2.40 FPKM, yes/no = 0/1). Multivariate Cox regression analysis showed that this gene score was the most important predictor of HCC (hazard ratio = 2.38, 95% confidence interval [CI] = 1.06-5.36, P = 0.036). Combining the gene score and fibrosis-4 index, a nomogram was constructed to predict the probability of HCC with an area under the receiver operating characteristic curve up to 0.950 (95% CI = 0.888-1.000, P = 7.0 × 10-5). Decision curve analysis revealed that the nomogram had a net benefit in HCC detection. The calibration curve showed that the nomogram had optimal concordance between the predicted and actual HCC probabilities. In conclusion, down-regulated posttreatment PBMC TAS1R3, ABCA3, and FOSL1 expression were significantly correlated with HCC development after HCV eradication. Decision-tree-based algorithms can refine the assessment of HCC risk for personalized HCC surveillance.