Intelligence-based medicine最新文献

Operation notes are a crucial component of patient care. However, writing them manually is prone to human error, particularly in high pressured clinical environments. Automatic generation of operation notes from video recordings can alleviate some of the administrative burdens, improve accuracy, and provide additional information. To achieve this for endoscopic pituitary surgery, 27-steps were identified via expert consensus. Then, for the 97-videos recorded for this study, a timestamp of each step was annotated by an expert surgeon. To automatically determine whether a step is present in a video, a three-stage architecture was created. Firstly, for each step, a convolution neural network was used for binary image classification on each frame of a video. Secondly, for each step, the binary frame classifications were passed to a discriminator for binary video classification. Thirdly, for each video, the binary video classifications were passed to an accumulator for multi-label step classification. The architecture was trained on 77-videos, and tested on 20-videos, where a 0.80 weighted-F₁ score was achieved. The classifications were inputted into a clinically based predefined template, and further enriched with additional video analytics. This work therefore demonstrates automatic generation of operative notes from surgical videos is feasible, and can assist surgeons during documentation.

A primary challenge for in vivo kidney magnetic resonance imaging (MRI) is the presence of different types of involuntary physiological motion, affecting the diagnostic utility of acquired images due to severe motion artifacts. Existing prospective and retrospective motion correction methods remain ineffective when dealing with complex large amplitude nonrigid motion artifacts. Here, we introduce an unsupervised deep learning-based image to image translation method between motion-affected and motion-free image domains, for correction of rigid-body, respiratory and nonrigid motion artifacts in vivo kidney MRI.

High resolution (i.e., 156 × 156 × 370 μm) ex vivo 3 Tesla MRI scans of 13 porcine kidneys (because of their anatomical homology to human kidney) were conducted using a 3D T2-weighted turbo spin echo sequence. Rigid-body, respiratory and nonrigid motion-affected images were then simulated using the magnitude-only ex vivo motion-free image set. Each 2D coronal slice of motion-affected and motion-free image volume was then divided into patches of 128 × 128 for training the model. We proposed to add normalised cross-correlation loss to cycle consistency generative adversarial network structure (NCC-CycleGAN), to enforce edge alignment between motion-corrected and motion-free image domains.

Our NCC-CycleGAN motion correction model demonstrated high performance with an in-tissue structural similarity index measure of 0.77 ± 0.08, peak signal-to-noise ratio of 26.67 ± 3.44 and learned perceptual image patch similarity of 0.173 ± 0.05 between the reconstructed motion-corrected and ground truth motion-free images. This corresponds to a significant respective average improvement of 34%, 23% and 39% (p < 0.05; paired t-test) for the three metrics to correct the three different types of simulated motion artifacts.

We demonstrated the feasibility of developing an unsupervised deep learning-based method for efficient automated retrospective kidney MRI motion correction, while preserving microscopic tissue structures in high resolution imaging.

The use of artificial intelligence (AI) programs in healthcare and medicine has steadily increased over the past decade. One major challenge affecting the use of AI programs is that the results of AI programs are sometimes not replicable, meaning that the performance of the AI program is substantially different in the external testing dataset when compared to its performance in the training or validation datasets. This often happens when the external testing dataset is very different from the training or validation datasets. Sex, ethnicity, and race are some of the most important biological and social determinants of health, and are important factors that may differ between training, validation, and external testing datasets, and may contribute to the lack of reproducibility of AI programs. We reviewed over 28,000 original research articles published in the three journals with the highest impact factors in each of 16 medical specialties between 2019 and 2022, to evaluate how often the sex, ethnic, and racial compositions of the datasets used to develop AI algorithms were reported. We also reviewed all currently used AI reporting guidelines, to evaluate which guidelines recommend specific reporting of sex, ethnicity, and race. We find that only 42.47 % (338/797) of articles reported sex, 1.4 % (12/831) reported ethnicity, and 7.3 % (61/831) reported race. When sex was reported, approximately 55.8 % of the study participants were female, and when ethnicity was reported, only 6.2 % of the study participants were Hispanic/Latino. When race was reported, only 29.4 % of study participants were non-White. Most AI guidelines (93.3 %; 14/15) also did not recommend reporting sex, ethnicity, and race. To have fair and ethnical AI, it is important that the sex, ethnic, and racial compositions of the datasets used to develop the AI program are known.

The scientific literature on depression detection using electroencephalogram (EEG) signals is extensive and offers numerous innovative approaches. However, these existing state-of-the-art (SOTA) have limitations that hinder their overall efficacy. They rely significantly on datasets with limited scope and accessibility, which introduces potential biases and diminishes generalizability. In addition, they concentrate on analyzing a single dataset, potentially overlooking the inherent variability and complexity of EEG patterns associated with depression. Moreover, certain SOTA methods employ deep learning architectures with exponential time complexity, resulting in computationally intensive and time-consuming training procedures. Therefore, their practicability and applicability in real-world scenarios are compromised. To address these limitations, a novel integrated methodology that combines the advantages of phase space reconstruction and deep neural networks is proposed. It employs publicly available EEG datasets, mitigating the inherent biases of exclusive data sources. Moreover, the method incorporates reconstructed phase space analysis, a feature engineering technique that captures more accurately the complex EEG patterns associated with depression. Simultaneously, the incorporation of a deep neural network component guarantees optimal efficiency and accurate, seamless classification. Using publicly available datasets, cross-dataset validation, and a novel combination of reconstructed phase space analysis and deep neural networks, the proposed method circumvents the shortcomings of current state-of-the-art (SOTA) approaches. This innovation represents a significant advance in enhancing the accuracy of depression detection and provides the base for EEG-based depression assessment tools applicable to real-world settings. The findings of the study provide a more robust and efficient model, which increases classification precision and decreases computing burden. The study findings layout the foundation for scalable, accessible mental health solutions, identification of the pathological deficits in affected brain tissues, and demonstrate the potential of technology-driven approaches to support and guide depressed individuals and enhance mental health outcomes.

Background

In military settings, determining whether a patient with abdominal pain requires emergency care can be challenging due to the absence or inexperience of medical staff. Misjudging the severity of abdominal pain can lead to delayed treatment or unnecessary transfers, both of which consume valuable resources. Therefore, our aim was to develop an artificial intelligence model capable of classifying the urgency of abdominal pain cases, taking into account patient characteristics.

Methods

We collected structured and unstructured data from patients with abdominal pain visiting South Korean military hospital emergency rooms between January 2015 and 2020. After excluding patients with missing values, 20,432 patients were enrolled. Structured data consisted of age, sex, vital signs, past medical history, and symptoms, while unstructured data included preprocessed free text descriptions of chief complaints and present illness. Patients were divided into training, validation, and test datasets in an 8:1:1 ratio. Using structured data, we developed four conventional machine learning models and a novel mixed model, which combined one of the best performing machine learning models with emergency medical knowledge. And we also created a deep learning model using both structured and unstructured data.

Results

Xgboost demonstrated the highest performance among the six models, with an area under the precision-recall curve (AUPRC) score of 0.61. The other five models achieved AUPRC scores as follows: logistic regression (0.24), decision tree (0.22), multi-layer perceptron (0.21), deep neural network (0.58), and mixed model (0.58).

Conclusion

This study is the first to develop an AI model for identifying emergency cases of abdominal pain in a military setting. With more balanced and better-structured datasets, clinically significant AI model could be developed based on the findings of this study.

Background

Electronic health records (EHR) were implemented to improve patient care, reduce healthcare disparities, engage patients and families, improve care coordination, and maintain privacy and security. Unfortunately, the mandated use of EHR has also resulted in significantly increased clerical and administrative burden, with physicians spending an estimated three-fourths of their daily time interacting with the EHR, which negatively affects within-clinic processes and contributes to burnout. In-room scribes are associated with improvement in all aspects of physician satisfaction and increased productivity, though less is known about the use of other technologies such as Google Glass (GG), Natural Language Processing (NLP) and Machine-Based Learning (MBL) systems. Given the need to decrease administrative burden on clinicians, particularly in the utilization of the EHR, there is a need to explore the intersection between varying degrees of technology in the clinical encounter and their ability to meet the aforementioned goals of the EHR.

Aims

The primary aim is to determine predictors of overall perception of care dependent on varying mechanisms used for documentation and medical decision-making in a routine clinical encounter. Secondary aims include comparing the perception of individual vignettes based on demographics of the participants and investigating any differences in perception questions by demographics of the participants.

Methods

Video vignettes were shown to 498 OhioHealth Physician Group patients and to ResearchMatch volunteers during a 15-month period following IRB approval. Data included a baseline survey to gather demographic and background characteristics and then a perceptual survey where patients rated the physician in the video on 5 facets using a 1 to 5 Likert scale. The analysis included summarizing data of all continuous and categorical variables as well as overall perceptions analyzed using multivariate linear regression with perception score as the outcome variable.

Results

Univariate modeling identified sex, education, and type of technology as three factors that were statistically significantly related to the overall perception score. Males had higher scores than females (p = 0.03) and those with lower education had higher scores (p < 0.001). In addition, the physician documenting outside of the room encounter had statistically significantly higher overall perception scores (mean = 22.2, p < 0.001) and the physician documenting in the room encounter had statistically significantly lower overall perception scores (mean = 15.3, p < 0.001) when compared to the other vignettes. Multivariable modeling identified all three of the univariably significant factors as independent factors related to overall perception score. Specifically, high school education had higher scores than associate/bachelor education (LSM = 21.6 vs.