Conventional severity-of-illness scoring systems have shown suboptimal performance for predicting in-intensive care unit (ICU) mortality in patients with severe pneumonia. This study aimed to develop and validate machine learning (ML) models for mortality prediction in patients with severe pneumonia. This retrospective study evaluated patients admitted to the ICU for severe pneumonia between January 2016 and December 2021. The predictive performance was analyzed by comparing the area under the receiver operating characteristic curve (AU-ROC) of ML models to that of conventional severity-of-illness scoring systems. Three ML models were evaluated: (1) logistic regression with L2 regularization, (2) gradient-boosted decision tree (LightGBM), and (3) multilayer perceptron (MLP). Among the 816 pneumonia patients included, 223 (27.3%) patients died. All ML models significantly outperformed the Simplified Acute Physiology Score II (AU-ROC: 0.650 [0.584–0.716] vs 0.820 [0.771–0.869] for logistic regression vs 0.827 [0.777–0.876] for LightGBM 0.838 [0.791–0.884] for MLP; P < 0.001). In the analysis for NRI, the LightGBM and MLP models showed superior reclassification compared with the logistic regression model in predicting in-ICU mortality in all length of stay in the ICU subgroups; all age subgroups; all subgroups with any APACHE II score, PaO₂/FiO₂ ratio < 200; all subgroups with or without history of respiratory disease; with or without history of CVA or dementia; treatment with mechanical ventilation, and use of inotropic agents. In conclusion, the ML models have excellent performance in predicting in-ICU mortality in patients with severe pneumonia. Moreover, this study highlights the potential advantages of selecting individual ML models for predicting in-ICU mortality in different subgroups.

According to recent studies, patients with COVID-19 have different feature characteristics on chest X-ray (CXR) than those with other lung diseases. This study aimed at evaluating the layer depths and degree of fine-tuning on transfer learning with a deep convolutional neural network (CNN)-based COVID-19 screening in CXR to identify efficient transfer learning strategies. The CXR images used in this study were collected from publicly available repositories, and the collected images were classified into three classes: COVID-19, pneumonia, and normal. To evaluate the effect of layer depths of the same CNN architecture, CNNs called VGG-16 and VGG-19 were used as backbone networks. Then, each backbone network was trained with different degrees of fine-tuning and comparatively evaluated. The experimental results showed the highest AUC value to be 0.950 concerning COVID-19 classification in the experimental group of a fine-tuned with only 2/5 blocks of the VGG16 backbone network. In conclusion, in the classification of medical images with a limited number of data, a deeper layer depth may not guarantee better results. In addition, even if the same pre-trained CNN architecture is used, an appropriate degree of fine-tuning can help to build an efficient deep learning model.

We aimed to develop a novel prediction model for early neurological deterioration (END) based on an interpretable machine learning (ML) algorithm for atrial fibrillation (AF)-related stroke and to evaluate the prediction accuracy and feature importance of ML models. Data from multicenter prospective stroke registries in South Korea were collected. After stepwise data preprocessing, we utilized logistic regression, support vector machine, extreme gradient boosting, light gradient boosting machine (LightGBM), and multilayer perceptron models. We used the Shapley additive explanation (SHAP) method to evaluate feature importance. Of the 3,213 stroke patients, the 2,363 who had arrived at the hospital within 24 h of symptom onset and had available information regarding END were included. Of these, 318 (13.5%) had END. The LightGBM model showed the highest area under the receiver operating characteristic curve (0.772; 95% confidence interval, 0.715–0.829). The feature importance analysis revealed that fasting glucose level and the National Institute of Health Stroke Scale score were the most influential factors. Among ML algorithms, the LightGBM model was particularly useful for predicting END, as it revealed new and diverse predictors. Additionally, the effects of the features on the predictive power of the model were individualized using the SHAP method.

Despite the significance of predicting the prognosis of idiopathic sudden sensorineural hearing loss (ISSNHL), no predictive models have been established. This study used artificial intelligence to develop prognosis models to predict recovery from ISSNHL. We retrospectively reviewed the medical data of 453 patients with ISSNHL (men, 220; women, 233; mean age, 50.3 years) who underwent treatment at a tertiary hospital between January 2021 and December 2019 and were followed up after 1 month. According to Siegel’s criteria, 203 patients recovered in 1 month. Demographic characteristics, clinical and laboratory data, and pure-tone audiometry were analyzed. Logistic regression (baseline), a support vector machine, extreme gradient boosting, a light gradient boosting machine, and multilayer perceptron were used. The outcomes were the area under the receiver operating characteristic curve (AUROC) primarily, area under the precision-recall curve, Brier score, balanced accuracy, and F1 score. The light gradient boosting machine model had the best AUROC and balanced accuracy. Together with multilayer perceptron, it was also significantly superior to logistic regression in terms of AUROC. Using the SHapley Additive exPlanation method, we found that the initial audiogram shape is the most important prognostic factor. Machine/deep learning methods were successfully established to predict the prognosis of ISSNHL.

Center of pressure (COP) during gait is a useful measure for assessing gait ability and has been investigated using platform or insole systems. However, these systems have inherent restrictions in repeated measure design or in obtaining true vertical force. This study proposes a novel method based on a pressure-sensitive mat system for COP measurement and presents normal reference values for the system. To explore repeatability, this work also investigated relative and absolute intra-rater reliabilities and determined the number of footfalls required to obtain a reliable measurement. Ninety healthy young adults participated and performed barefoot walking on a force-sensitive mat at a comfortable and fast pace. The time points and subphase duration of the stance phase, displacement ranges, and mean locations of COP and velocity of COP excursion were parameterized. The results showed acceptable and consistent variabilities of the parameters. Seven footfalls were determined as the threshold for most parameters to show a good to reasonable level of reliability. In conclusion, the presented method can be used as a reliable measurement for COP excursion, and it is recommended that more than seven footfalls be collected to ensure a high level of reliability.

Kim DH, et al. Korean J Radiol. 2024 Feb;25:e27. https://doi.org/10.3348/kjr.2023.0671