Kim Y, et al. J Korean Med Sci. 2021 Jul;36(31):e198. https://doi.org/10.3346/jkms.2021.36.e198

Background: The prediction of successful weaning from mechanical ventilation (MV) in advance of intubation can facilitate discussions regarding end-of-life care before unnecessary intubation. Objective: We aimed to develop a machine learning–based model that predicts successful weaning from ventilator support based on routine clinical and laboratory data taken before or immediately after intubation. Methods: We used the Medical Information Mart for Intensive Care IV database, which is an open-access database covering 524,740 admissions of 382,278 patients in Beth Israel Deaconess Medical Center, United States, from 2008 to 2019. We selected adult patients who underwent MV in the intensive care unit (ICU). Clinical and laboratory variables that are considered relevant to the prognosis of the patient in the ICU were selected. Data collected before or within 24 hours of intubation were used to develop machine learning models that predict the probability of successful weaning within 14 days of ventilator support. Developed models were integrated into an ensemble model. Performance metrics were calculated by 5-fold cross-validation for each model, and a permutation feature importance and Shapley additive explanations analysis was conducted to better understand the impacts of individual variables on outcome prediction. Results: Of the 23,242 patients, 19,025 (81.9%) patients were successfully weaned from MV within 14 days. Using the preselected 46 clinical and laboratory variables, the area under the receiver operating characteristic curve of CatBoost classifier, random forest classifier, and regularized logistic regression classifier models were 0.860 (95% CI 0.852-0.868), 0.855 (95% CI 0.848-0.863), and 0.823 (95% CI 0.813-0.832), respectively. Using the ensemble voting classifier using the 3 models above, the final model revealed the area under the receiver operating characteristic curve of 0.861 (95% CI 0.853-0.869), which was significantly better than that of Simplified Acute Physiology Score II (0.749, 95% CI 0.742-0.756) and Sequential Organ Failure Assessment (0.588, 95% CI 0.566-0.609). The top features included lactate and anion gap. The model’s performance achieved a plateau with approximately the top 21 variables. Conclusions: We developed machine learning algorithms that can predict successful weaning from MV in advance to intubation in the ICU. Our models can aid the appropriate management for patients who hesitate to decide on ventilator support or meaningless end-of-life care.

Background: This study aimed to develop a machine learning-based 2-year risk prediction model for early identification of patients with rapid progressive immunoglobulin A nephropathy (IgAN). We also assessed the model’s performance to predict the long-term kidney-related outcome of patients. Methods: A retrospective cohort of 1,301 patients with biopsy-proven IgAN from two tertiary hospitals was used to derive and externally validate a random forest-based prediction model predicting primary outcome (30% decline in estimated glomerular filtration rate from baseline or end-stage kidney disease requiring renal replacement therapy) and secondary outcome (improvement of proteinuria) within 2 years after kidney biopsy. Results: For the 2-year prediction of primary outcomes, precision, recall, area-under-the-curve, precision-recall-curve, F1, and Brier score were 0.259, 0.875, 0.771, 0.242, 0.400, and 0.309, respectively. The values for the secondary outcome were 0.904, 0.971, 0.694, 0.903, 0.955, and 0.113, respectively. From Shapley Additive exPlanations analysis, the most informative feature identifying both outcomes was baseline proteinuria. When Kaplan-Meier analysis for 10-year kidney outcome risk was performed with three groups by predicting probabilities derived from the 2-year primary outcome prediction model (low, moderate, and high), high (hazard ratio [HR], 13.00; 95% confidence interval [CI], 9.52–17.77) and moderate (HR, 12.90; 95% CI, 9.92–16.76) groups showed higher risks compared with the low group. From the 2-year secondary outcome prediction model, low (HR, 1.66; 95% CI, 1.42–1.95) and moderate (HR, 1.42; 95% CI, 0.99–2.03) groups were at greater risk for 10-year prognosis than the high group. Conclusion: Our machine learning-based 2-year risk prediction models for the progression of IgAN showed reliable performance and effectively predicted long-term kidney outcome.

Background: Pulmonary tuberculosis (PTB) poses a global health challenge owing to the time-intensive nature of traditional diagnostic tests such as smear and culture tests, which can require hours to weeks to yield results. Objective: This study aimed to use artificial intelligence (AI)–based chest radiography (CXR) to evaluate the infectivity of patients with PTB more quickly and accurately compared with traditional methods such as smear and culture tests. Methods: We used DenseNet121 and visualization techniques such as gradient-weighted class activation mapping and local interpretable model-agnostic explanations to demonstrate the decision-making process of the model. We analyzed 36,142 CXR images of 4492 patients with PTB obtained from Severance Hospital, focusing specifically on the lung region through segmentation and cropping with TransUNet. We used data from 2004 to 2020 to train the model, data from 2021 for testing, and data from 2022 to 2023 for internal validation. In addition, we used 1978 CXR images of 299 patients with PTB obtained from Yongin Severance Hospital for external validation. Results: In the internal validation, the model achieved an accuracy of 73.27%, an area under the receiver operating characteristic curve of 0.79, and an area under the precision-recall curve of 0.77. In the external validation, it exhibited an accuracy of 70.29%, an area under the receiver operating characteristic curve of 0.77, and an area under the precision-recall curve of 0.8. In addition, gradient-weighted class activation mapping and local interpretable model-agnostic explanations provided insights into the decision-making process of the AI model. Conclusions: This proposed AI tool offers a rapid and accurate alternative for evaluating PTB infectivity through CXR, with significant implications for enhancing screening efficiency by evaluating infectivity before sputum test results in clinical settings, compared with traditional smear and culture tests.

Hospital falls are the most prevalent adverse event in healthcare, posing significant risks to patient health outcomes and institutional care quality. The effectiveness of several fall prediction models currently in use is limited by various clinical factors. This study explored the efficacy of merging real-time location system (RTLS) data with clinical information to enhance the accuracy of in-hospital fall predictions. The model performances were compared based on the clinical data, RTLS data, and a hybrid approach using various evaluation metrics. The RTLS and integrated clinical data were obtained from 22,201 patients between March 2020 and June 2022. From the initial cohort, 118 patients with falls and 443 patients without falls were included. Predictive models were developed using the XGBoost algorithm across three distinct frameworks: clinical model, RTLS model, and clinical + RTLS model. The model performance was evaluated using metrics, such as AUROC, AUPRC, accuracy, PPV, sensitivity, specificity, and F1 score. Shapley additive explanation values were used to enhance the model interpretability. The clinical model yielded an AUROC of 0.813 and AUPRC of 0.407. The RTLS model demonstrated superior fall prediction capabilities, with an AUROC of 0.842 and AUPRC of 0.480. The clinical + RTLS model excelled further, achieving an AUROC of 0.853 and AUPRC of 0.497. Feature importance analysis revealed that movement patterns of patients on the last day of their stay were significantly associated with falls, together with elevated RDW levels, sedative administration, age. This study underscored the advantages of combining RTLS data with clinical information to predict in-hospital falls more accurately. This innovative technology-driven approach may enhance early fall risk detection during hospitalization, potentially preventing falls, improving patient safety, and contributing to more efficient healthcare delivery.

Background: Dementia is a growing health burden worldwide; however, scalable, low-cost strategies for early risk identification are lacking. Because cardiac and cognitive aging are closely associated, a routinely collected 12-lead electrocardiogram (ECG), interpreted using foundation model embeddings, is a promising and accessible candidate marker. However, it is unclear whether such embeddings carry predictive information independent of chronological age—the dominant risk factor for dementia—and whether the underlying changes follow a detectable prediagnostic trajectory. Objective: This study aimed to investigate whether foundation model embeddings of routine ECGs carry information that distinguishes patients later diagnosed with cognitive impairment or dementia from matched controls, beyond chronological age, and, if so, whether the underlying ECG changes intensify as diagnosis approaches. Methods: In a retrospective two-site Korean cohort (Severance Hospital, development n=52,355; Yongin Severance Hospital, external n=6500), embeddings from the knowledge-enhanced diagnosis foundation model supported an ECG-only classifier that discriminated patients diagnosed with cognitive impairment or dementia within the following 10 years from matched controls, using only ECGs interpreted as normal sinus rhythm. We compared an age-only model with a late fusion of age and ECG-derived probability and tested whether age-residualized embedding is predictive as direct evidence of age-independent information. Discrimination was summarized by the area under the receiver operating characteristic curve (AUROC) with 95% bootstrap CIs, and DeLong tests and net reclassification improvement (NRI) were used for between-model comparisons. To trace the prediagnostic time course, we analyzed ECG phenotype labels across four windows before diagnosis in an exact-age-matched case-control set. Results: Chronological age strongly discriminated cases from controls in both cohorts (external AUROC 0.809, 95% CI 0.799-0.819). Adding ECG-derived probability improved discrimination, calibration, and net benefit over age alone (external AUROC 0.825, 95% CI 0.816-0.834; DeLong P<.001; positive NRI). Critically, even after removing the age-predictable component, age-residualized embedding retained modest but reproducible discrimination in both internal (AUROC 0.590, 95% CI 0.581-0.599) and external (AUROC 0.579, 95% CI 0.567-0.592) validation (both P<.001), indicating dementia-relevant information that was not explained by chronological age. Of the 105 ECG labels, 19 (18.1%) showed a case-control difference that widened significantly toward diagnosis, dominated by nonspecific ST segment and T wave changes, acute ischemic patterns (acute myocardial infarction), and atrial and junctional arrhythmias (atrial flutter, ectopic atrial, and accelerated junctional rhythms), with divergence emerging mainly in the proximal (1-3-year) and imminent (0-1-year) windows. Conclusions: Routine ECG foundation model embeddings encode aging-related electrophysiological information that is complementary to and not fully explained by chronological age and intensifies as diagnosis approaches, with concurrent shifts toward ischemic and arrhythmic ECG patterns whose mechanism cannot be determined from these data. These findings suggest that routine ECG is a low-cost, age-complementary triage signal for cognitive impairment and dementia screening, rather than a stand-alone diagnostic test.