Intelligence-based medicine最新文献

The depth and breadth of data produced in the modern cardiac intensive care unit (CICU) poses challenges to clinicians and researchers. Artificial intelligence (AI) and machine learning (ML) methodologies have been increasingly used to provide insights into this complex patient population. Major analytical tasks where ML methodology can be applied in the CICU and other critical care settings include mortality risk stratification, prognostication, non-fatal event prediction, diagnosis, phenotyping, identification of occult heart disease from the electrocardiogram and interpretation of echocardiographic images. In this review, we will discuss existing and future applications of different ML methods for CICU and other critical care populations, including penalized regression, standard ML methods (e.g., tree-based and other non-linear approaches) and advanced ML methods (e.g., deep learning and neural networks). While comparatively few published studies have applied ML methods in CICU populations, a more robust literature including patients with acute cardiovascular disease and non-cardiovascular critical illness can provide insights into CICU care. The CICU of the future is likely to utilize a sophisticated array of ML algorithms to streamline patient care by facilitating early recognition, diagnosis, phenotyping, and intervention for critically ill or deteriorating patients to improve providers’ cognitive load.

Objectives

To evaluate the similarities in microphotographic images across different smartphones and to establish whether computer vision can use microphotographs to successfully classify dental caries.

Method

A universal clip-type microscope with 60x optical zoom was selected to perform in vitro microphotography of extracted teeth. For the first objective, areas of cariogenic interest were physically labelled by dentists and eight smartphones were used to capture images of tooth decays with the microscope fitted over the primary camera lens. For the second objective, 233 microphotography images were acquired and virtually augmented to produce 1631 images that were categorized digitally by an international caries classification system for computer vision-based object detection (YOLO.v4). Five practitioners independently labelled randomly selected images from the test dataset following the caries classification system which were subsequently used to evaluate the diagnostic test accuracy of the YOLO model.

Result

A significant overall mean square error [F (df) = 4.03 (6); P < 0.05] was observed while Bhattacharya's distance evaluation produced no significant differences [F (df) = 1.60 (6); P > 0.05] across all eight smartphone derived datasets. Index and reference test comparisons determined an overall sensitivity of 0.99 and specificity of 0.94 for the trained YOLO.v4 and highly significant correlations (r > 0.9, P < 0.001) to the classifications labelled by the dental practitioners.

Conclusion

Non-standardized images of tooth caries captured by different smartphones generated an accurate diagnostic model for classifying carious lesions that was similar to the visual assessments performed by experienced dental practitioners.

Background

Super-utilizers consume the greatest share of resource intensive healthcare (RIHC) and reducing their utilization remains a crucial challenge to healthcare systems in the United States (U.S.). The objective of this study was to predict RIHC among U.S. counties, using routinely collected data from the U.S. government, including information on consumer spending, offering an alternative method for identifying super-utilization among population units rather than individuals.

Methods

Cross-sectional data from 5 governmental sources in 2017 were used in a machine learning pipeline, where target-prediction features were selected and used in 4 distinct algorithms. Outcome metrics of RIHC utilization came from the American Hospital Association and included yearly: (1) emergency rooms visit, (2) inpatient days, and (3) hospital expenditures. Target-prediction features included: 149 demographic characteristics from the U.S. Census Bureau, 151 adult and child health characteristics from the Centers for Disease Control and Prevention, 151 community characteristics from the American Community Survey, and 571 consumer expenditures from the Bureau of Labor Statistics. SHAP analysis identified important target-prediction features for 3 RIHC outcome metrics.

Results

2475 counties with emergency rooms and 2491 counties with hospitals were included. The median yearly emergency room visits per capita was 0.450 [IQR:0.318, 0.618], the median inpatient days per capita was 0.368 [IQR: 0.176, 0.826], and the median hospital expenditures per capita was $2104 [IQR: $1299.93, 3362.97]. The coefficient of determination (R²), calculated on the test set, ranged between 0.267 and 0.447. Demographic and community characteristics were among the important predictors for all 3 RIHC outcome metrics.

Conclusions

Integrating diverse population characteristics from numerous governmental sources, we predicted 3-outcome metrics of RIHC among U.S. counties with good performance, offering a novel and actionable tool for identifying super-utilizer segments in the population. Wider integration of routinely collected data can be used to develop alternative methods for predicting RIHC among population units.

In this proposed research work, a computerized Hybrid U-Net model for supplying colon glandular morphometric and cancer grade information is demonstrated. The solution is put forth by incorporating three distinctive structural elements—Advanced Convolutional Learning Modules, Attention Modules, and Multi-Scalar Transitional Modules—into the conventional U-Net architecture. By combining these modules, complex multi-level convolutional feature learning further encompassed with target-specified attention and increased effective receptive-field-size are produced. Three publicly accessible datasets—CRAG, GlaS challenge, LC-25000 dataset, and an internal, proprietary dataset Hospital Colon (HosC)—are used in experiments. The suggested model also produced competitive results for the gland detection and segmentation task in terms of Object-Dice Index as ((0.950 for CRAG), (GlaS: (0.951 for Test A & 0.902 for Test B)), (0.954 for LC-25000), (0.920 for HosC)), F1-score as ((0.921 for CRAG), (GlaS: (0.945 for Test A & 0.923 for Test B)), (0.913 for LC-25000), (0.955 for HosC)), and Object-Hausdorff Distance ((90.43 for CRAG), (GlaS: (23.11 for Test A & 71.47 for Test B)), (96.24 for LC-25000), (85.41 for HosC)). Pathologists evaluated the generated segmented glandular areas and assigned a mean score as ((9.25 for CRAG), (GlaS: (9.32 for Test A & 9.28 for Test B)), (9.12 for LC-25000) (9.14 for HosC)). The proposed model successfully completed the task of determining the cancer grade with the following results: Precision as ((0.9689 for CRAG), (0.9721 for GlaS), (1.0 for LC-25000), (1.0 for HosC)), Specificity (0.8895 for CRAG), (0.9710 for GlaS), (1.0 for LC-25000), (1.0 for HosC)), and Sensitivity ((0.9677 for CRAG), (0.9722 for GlaS), (0.9995 for LC-25000), (0.9932 for HosC)). Additionally, the Gradient-Weighted class activation mappings are provided to highlight the critical regions that the suggested model believes are essential for accurately predicting cancer. These visualizations are further reviewed by skilled pathologists and assigned with the mean scores as ((9.37 for CRAG), (9.29 for GlaS), (9.09 for LC-25000), and (9.91 for HosC)). By offering a referential opinion during the morphological assessment and diagnosis formulation in histopathology images, these results will help the pathologists and contribute towards reducing inadvertent human mistake and accelerating the cancer detection procedure.

Cerebrovascular disease is one of the world's leading causes of death. Blood vessel segmentation is a primary stage in diagnosing. Although a few deep neural networks have been suggested to automate volumetric brain blood vessel segmentation, few studies have considered the relevance of the evaluation metrics to diagnosing cerebrovascular disease due to the complicated nature of this task. This study aimed to understand if brain vasculature segmentation using a convolutional neural network (CNN) could meet radiologists' requirements for disease diagnosis. We employed a deeply supervised attention-gated 3D U-Net trained based on the Focal Tversky loss function to extract brain vasculatures from volumetric magnetic resonance angiography (MRA) images. Here we show that our training procedure led to biologically relevant results despite not scoring well using the Dice score, a common metric for algorithm evaluation. We achieved Dice (±SD) = 0.71 ± 0.02 and two radiologists confirmed and validated that our method successfully captured the major blood vessel branches of the circle of Willis (CoW) having biological importance, including internal carotid artery (ICA), middle cerebral artery (MCA), anterior cerebral artery (ACA), and posterior cerebral artery (PCA). Adding radiologists' expert opinions, we could fill this gap that using only the current common evaluation metrics, such as the Dice coefficient, is not enough for brain vessel segmentation assessment. These results suggest the additional value for computational approaches that are designed with end-user stakeholders in mind.

Background

Hospital length of stay (LoS) varies widely across hip (HA) and knee arthroplasty (KA) patients and depends on multiple factors. Prediction methods are necessary to improve hospital capacity planning and identify patients at risk of long LoS. This study aims (1) to compare the performance of previously applied machine learning (ML) as well as regression methods for either LoS classification or regression in a multi-hospital setting for primary HA and KA patients. In addition, the study aims (2a) to assess which variables are the most important predictors for LoS prediction and, specifically, (2b) whether patient-reported outcome measures (PROMs) collected before surgery act as important predictors.

Methods

2611 primary HA and 2077 primary KA patients from eight German hospitals were included to train and test extreme gradient boosting (XGBoost), naïve Bayes (NB) and logistic regression (LogReg) for classification, and XGBoost as well as a linear regression (LinReg) for regression. Area under the receiver operating characteristics curve (AUC) and mean absolute error (MAE) were used as primary performance indicators for classification and regression.

Results

For classification, the highest AUC was reached by XGBoost and LogReg (AUC = 0.81) in the HA sample, whereas NB was statistically significantly outperformed by both other methods. In the KA sample, no statistical difference between any method was found, and AUC was lower for all models compared with HA. For regression, MAE was lowest for XGBoost (1.43 days for HA and 1.21 days for KA). PROMs and hospital indicators were among the most relevant predictors in all cases.

Conclusion

The study demonstrated robust performance of ML in predicting LoS. PROMs reflect relevant features for prediction. They should be routinely collected and used for practical applications. XGBoost may act as a superior prediction tool compared to regression or other ML models in certain circumstances.

With the burgeoning population of patients with cognitive disorders such as dementia, healthcare is already having significant difficulty in caring for this new population of patients. However, with the last decade's advances in neural networks, it is possible to begin creating software which may aid in healthcare for these patients. Specifically, it may aid in diagnosis, such as in expediting cognitive examinations. Within this paper, we describe a custom neural network utilizing a SqueezeNet which is used to classify a custom dataset of hand-drawn images commonly used in cognitive examinations. We demonstrate that our model has 97% accuracy. Specifically, this enables the development of entire automated and accurate cognitive examinations. The work presented here demonstrates neural networks may assist with healthcare for patients with cognitive disorders, having impact upon the fields of neurology, psychiatry, and family medicine. Importantly, within the context of the COVID-19 pandemic restricting in-person visits and promoting telemedicine, this provides the foundations to transition cognitive examinations to a telemedicine modality.

The rapid progress of artificial intelligence (AI) and the adoption of Large Language Models (LLMs) suggests that these technologies will transform healthcare in the coming years. We present a primer on LLMs for clinicians, focusing on OpenAI's Generative Pretrained Transformer-4 (GPT-4) model which powers ChatGPT as a use-case, as it has already seen record-breaking uptake in usage. ChatGPT generates natural-sounding text based on patterns observed from vast amounts of training data. The core strengths of ChatGPT and LLMs in healthcare applications include summarization and text generation, rapid adaptation and learning, and ease of customization and integration into existing applications. However, clinicians should also recognize the limitations of LLMs, most notably concerns about inaccuracy, privacy, accountability, transparency, and explainability. Clinicians must embrace the opportunity to explore, engage, and lead in the responsible integration of LLMs, harnessing their potential to revolutionize patient care and drive advancements in an ever-evolving healthcare landscape.

To perform data mining projects for knowledge discovery based on health data produced in a daily health care stored in electronic health records (EHR) can be time consuming. This study exemplifies that the involvement of a data scientist improves classification performances. We have performed a case study that comprises two real world medical research projects, comparing feature engineering and knowledge discovery based on classification performance. Project (P1) comprised 82,742 patients with the research question “Can we predict patient falls by use of EHR data” and the second project (P2) included 23,396 patients with the focus on “Negative side effects of antiepileptic drug consumption on bone structure”.

The results concluded three salient results. (i) It is valuable for medical researchers to involve a data scientist when medical research based on real world medical data is performed. The findings were justified with an analysis of classification metrics when iteratively engineered features were used. The features were generated from domain experts and computer scientists in collaboration with medical researchers. We gave this process the name domain knowledge-driven feature engineering (KDFE).

To evaluate the classification performance the metric area under the receiver operating characteristic curve (AUROC) was used. (ii) Domain experts are benefited in quantitative terms by KDFE. When KDFE was compared to baseline, the average classification performance measured by AUROC for the engineered features rose for P1 from 0.62 to 0.82 and for P2 from 0.61 to 0.89 (p-values << 0.001). (iii) The engineered features were represented in a systematic structure, which is the foundation of a theoretical model for automated KDFE (aKDFE).

To our knowledge, this is the first study that proves that via quantitative measures KDFE adds value to real-world. However, the method is not limited to the medical domain. Other areas with similar data properties should also benefit from KDFE.

In a pharmaceutical context, metabolomics is an underexplored area of research. Nevertheless, its utility in clinical pathology, biomarker discovery, metabolic subtyping, and prognosis has transformed medicine. As this young domain evolves, its promise as an approach to drug discovery becomes more evident. It has established links between human phenotypes and quantitative biochemical parameters, enabling the construction of genome-scale metabolic networks. While the human metabolome is too vast for manual analysis, machine learning (ML) algorithms can efficiently recognize latent patterns in complex, large sets of metabolic data. ML-driven studies of the human metabolome and its constituents can inform efforts to reduce the quantity of resources spent at critical stages of the pipeline by facilitating target identification, mechanism of action elucidation, lead discovery, off-target effect evaluation, and in vivo response prediction. Metabolism-informed ML models generate insights that significantly advance efforts to reduce attrition rates and optimize drug efficacy. While applications of more advanced ML methods in studies of human metabolism are just beginning to form a body of literature, they have yielded promising results with implications for data-driven drug discovery.