Low birth weight (LBW) is an important adverse neonatal outcome and a public health concern. This study aimed to develop and validate an LBW risk prediction model based on machine learning algorithms using routine clinical indicators from early and mid-pregnancy. This single-center retrospective cohort study included 4,498 singleton pregnant women from a tertiary hospital in Urumqi (February 2014 – December 2024), among whom 579 (12.9
Whilst the use of artificial intelligence (AI) in healthcare is growing, with promises that AI will reduce costs and improve efficiency, there is also concern that AI systems can perpetuate health inequalities. Prior work has explored patients’ trust in AI, but less attention has been given to how patients conceptualise the fairness of AI decision-making in healthcare and which factors influence these views. We conducted a cross-sectional survey of UK patients, recruited through public and patient involvement and engagement (PPIE) networks, exploring perceptions of fairness across two AI vignettes: a patient-facing symptom checker and a clinician-facing cancer treatment decision-support tool. Respondents rated the perceived impact of 13 factors on the fairness of AI decisions and completed free-text responses to identify any additional factors. Demographics and survey variables were reported using descriptive statistics, along with Pearson’s chi-square test for differences in attitudes. Associations were estimated using ordinal logistic regression. Free-text responses were analysed thematically. A total of 149 respondents from the UK completed the survey. The factor perceived to most strongly undermine fairness of AI decisions was the absence of human supervision, with 89.9
Prompt diagnosis and timely reperfusion are essential in ST-segment elevation myocardial infarction (STEMI). Large language models, such as ChatGPT, may assist non-specialist providers by rapidly interpreting electrocardiograms (ECGs) and recommending evidence-based treatment, particularly in high-volume or resource-limited settings. To assess ChatGPT’s accuracy in localizing STEMI (anterior, inferior, lateral, or true left bundle branch block [LBBB]) on 12-lead ECGs, and guiding fibrinolysis vs. primary percutaneous coronary intervention (PCI) based on estimated travel times to a tertiary center. In this retrospective single-center study, 210 STEMI patients who underwent coronary angiography were included. For each case, anonymized ECG data and the location where the ECG was recorded were input into ChatGPT. Using GPS-based travel estimates to Konya City Hospital, ChatGPT recommended fibrinolysis if transit exceeded 120 min, otherwise PCI. Diagnostic accuracy was compared with cardiologist-confirmed STEMI localization, and agreement with real-life treatment decisions was evaluated via Cohen’s kappa. ChatGPT correctly identified 192 of 210 STEMIs (91.4
Early prognostic models for critically ill patients with acute myocardial infarction (AMI) often use first-day or full-stay data, and the incremental value of nonlinear machine learning over transparent regression remains uncertain. We reconstructed a retrospective true 6-hour landmark cohort in the Medical Information Mart for Intensive Care IV (MIMIC-IV) for development and internal testing, followed by one correction-specific, one-time external evaluation in the eICU Collaborative Research Database (eICU-CRD) after model locking. Eligible adults had current-admission AMI and were alive and under intensive care unit (ICU) observation at 6 h. The endpoint was index-hospital death after the landmark and within 28 days of ICU admission. We compared a first-available baseline logistic model, a parsimonious logistic model, and compact 0-6-hour logistic and LightGBM models. The compact models used the same 14 predictors. We assessed discrimination, Brier score, calibration, paired differences, and decision curves. The MIMIC risk set contained 8,142 admissions and 1,200 events; the untouched internal test set contained 2,436 admissions and 359 events. The eICU risk set contained 10,435 admissions and 799 events. In internal testing, compact logistic regression and LightGBM achieved areas under the receiver operating characteristic curve (AUROCs) of 0.785 (95
To investigate the dynamic trajectory of caregiver burden among family caregivers of polytrauma patients at different post-discharge time points, identify key determinants, and explore factors associated with caregiver burden using machine learning methods. This longitudinal study enrolled 275 family caregivers from a trauma center using convenience sampling. Caregiver burden was assessed using the Caregiver Burden Inventory (CBI) at 1 week, 1 month, 3 months, and 6 months post-discharge. Concurrent data collection included patients’ clinical indicators, caregivers’ demographic characteristics, and psychosocial metrics. Candidate variables were initially screened through Pearson correlation analysis (|r| > 0.3), followed by multicollinearity assessment, with subsequent application of eXtreme Gradient Boosting (XGBoost) and Random Forest (RF) models for exploratory analysis. Model fit was evaluated using the coefficient of determination (R²). The CBI showed a continuous downward trend: 37.90 ± 15.80 (1 week), 31.60 ± 17.46 (1 month), 23.43 ± 17.78 (3 months), and 16.45 ± 17.88 (6 months), and the differences were statistically significant at different time points (P < 0.01). The change trend of the scores of each dimension of the CBI was consistent with the total score. In the exploratory machine learning analysis, the XGBoost model showed the highest model fit (R² = 0.99). Features contributing prominently to the model included the degree of left lower limb injury and the scores of each dimension of caregiver burden. Given that the CBI total score is mathematically derived from its subscale scores, the high model fit should be interpreted cautiously and should not be considered evidence of prospective predictive accuracy. Family caregiver burden was gradually relieved with the rehabilitation of patients. Early subjective burden perception and patient injury were associated with caregiver burden. The exploratory machine learning analysis provided additional insight into the relative contributions of factors associated with caregiver burden; however, the observed model performance should not be interpreted as evidence of clinical predictive utility.
