Xiaokai Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Xiaokai Wang has a total of 36 co-authors. Here are the top co-authors by number of shared publications:
Aimin Wang
Department of Emergency, Xiangya Hospital, Central South University, Changsha 410008, Hunan Province, China
1 shared publication- Baorui Huang
Department of Emergency, Xiangya Hospital, Central South University, Changsha 410008, Hunan Province, China
1 shared publication - Ben Liu
Emergency Department, Xiangya Hospital, Central South University, Changsha, Hunan, China
1 shared publication - Caiyun He
Hunan Normal University
1 shared publication - Changyun Fang
Department of Stomatology, Xiangya Hospital, Central South University, Changsha, China
1 shared publication Guoqing Huang
1 shared publication
Latest Publications
- Open Access
- 24 September 2026
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 28, e98205
Clinical Surveillance Technologies in Nonintensive Care Unit Hospital Settings: Systematic Review and Bayesian Network Meta-Analysis of Randomized Trials
Background: Failure to recognize clinical deterioration in hospitalized patients has prompted the development of rule-based electronic surveillance (RB-ES), predictive model–based electronic surveillance (PM-ES), and continuous physiologic monitoring (CPM). However, their comparative effects on patient-centered outcomes remain uncertain. Objective: The aim of this study is to compare the clinical effects of emergent intensive RB-ES, PM-ES, CPM, and local standard care in nonintensive care unit hospital settings. Methods: We searched PubMed, Embase via online, the Cochrane Central Register of Controlled Trials via the Cochrane Library, and the Web of Science Core Collection from inception through February 28, 2026. Targeted supplementary surveillance, including trial-registry follow-up, backward and forward citation tracking, and known-item searches for newly available reports of registered trials, continued through July 22, 2026. Randomized, cluster-randomized, randomized crossover, and stepped-wedge trials were eligible. Interventions were classified according to their principal randomized function rather than their commercial or algorithmic labels. The primary outcomes were all-cause in-hospital or up-to-30-day mortality, and unplanned or emergent intensive care unit transfers. Bayesian random-effects network meta-analyses were performed using study-level adjusted relative effects. Secondary outcomes were evaluated using construct-specific pairwise meta-analyses. Risk of bias was assessed using the appropriate RoB 2 tool, and confidence in the network estimates was manually evaluated using the CINeMA (Confidence in Network Meta-Analysis) framework. Trial sequential analysis was retained as an exploratory supplementary analysis. Results: The review included 28 independent trials. Nine met strict digital-surveillance criteria; 7 contributed to at least 1 network, while 2 contributed only to sensitivity analyses because of outcome-definition or zero-event limitations. Six trials with 13,716 observations contributed to the mortality network. Compared with standard care, odds ratios were 0.91 (95% credible interval [CrI] 0.35‐2.30) for RB-ES, 1.22 (95% CrI 0.61‐2.31) for PM-ES, and 0.70 (95% CrI 0.36‐1.29) for CPM. Six trials with 13,441 observations contributed to the intensive care unit transfer network. Corresponding odds ratios were 1.20 (95% CrI 0.56‐2.59), 0.95 (95% CrI 0.58‐1.52), and 0.70 (95% CrI 0.37‐1.28), respectively. Expanded rule-based surveillance-response systems did not clearly reduce cardiac arrest or cardiopulmonary resuscitation (OR 0.94, 95% CI 0.77‐1.14). CPM (ratio of means 0.91, 95% CI 0.76‐1.09) and predictive model–based surveillance (ratio of means 0.99, 95% CI 0.47‐2.08) showed no clear effect on hospital length of stay. Trial sequential analyses were inconclusive. Confidence in all network comparisons was very low. Conclusions: Current randomized evidence does not establish a reliable clinical-effectiveness hierarchy among RB-ES, PM-ES, and CPM. Continuous monitoring showed directionally favorable but imprecise estimates for several outcomes. Surveillance strategies should therefore not be selected solely according to algorithm class or predictive complexity; their clinical effects may also depend on the target population, background monitoring, workflow integration, alert presentation, and clinical response pathway.
- 24 September 2026
- Published by JMIR Publications Inc. in Journal of Medical Internet Research
- Vol. 28, e98205
Open AccessClinical Surveillance Technologies in Nonintensive Care Unit Hospital Settings: Systematic Review and Bayesian Network Meta-Analysis of Randomized Trials
Background: Failure to recognize clinical deterioration in hospitalized patients has prompted the development of rule-based electronic surveillance (RB-ES), predictive model–based electronic surveillance (PM-ES), and continuous physiologic monitoring (CPM). However, their comparative effects on patient-centered outcomes remain uncertain. Objective: The aim of this study is to compare the clinical effects of emergent intensive RB-ES, PM-ES, CPM, and local standard care in nonintensive care unit hospital settings. Methods: We searched PubMed, Embase via online, the Cochrane Central Register of Controlled Trials via the Cochrane Library, and the Web of Science Core Collection from inception through February 28, 2026. Targeted supplementary surveillance, including trial-registry follow-up, backward and forward citation tracking, and known-item searches for newly available reports of registered trials, continued through July 22, 2026. Randomized, cluster-randomized, randomized crossover, and stepped-wedge trials were eligible. Interventions were classified according to their principal randomized function rather than their commercial or algorithmic labels. The primary outcomes were all-cause in-hospital or up-to-30-day mortality, and unplanned or emergent intensive care unit transfers. Bayesian random-effects network meta-analyses were performed using study-level adjusted relative effects. Secondary outcomes were evaluated using construct-specific pairwise meta-analyses. Risk of bias was assessed using the appropriate RoB 2 tool, and confidence in the network estimates was manually evaluated using the CINeMA (Confidence in Network Meta-Analysis) framework. Trial sequential analysis was retained as an exploratory supplementary analysis. Results: The review included 28 independent trials. Nine met strict digital-surveillance criteria; 7 contributed to at least 1 network, while 2 contributed only to sensitivity analyses because of outcome-definition or zero-event limitations. Six trials with 13,716 observations contributed to the mortality network. Compared with standard care, odds ratios were 0.91 (95% credible interval [CrI] 0.35‐2.30) for RB-ES, 1.22 (95% CrI 0.61‐2.31) for PM-ES, and 0.70 (95% CrI 0.36‐1.29) for CPM. Six trials with 13,441 observations contributed to the intensive care unit transfer network. Corresponding odds ratios were 1.20 (95% CrI 0.56‐2.59), 0.95 (95% CrI 0.58‐1.52), and 0.70 (95% CrI 0.37‐1.28), respectively. Expanded rule-based surveillance-response systems did not clearly reduce cardiac arrest or cardiopulmonary resuscitation (OR 0.94, 95% CI 0.77‐1.14). CPM (ratio of means 0.91, 95% CI 0.76‐1.09) and predictive model–based surveillance (ratio of means 0.99, 95% CI 0.47‐2.08) showed no clear effect on hospital length of stay. Trial sequential analyses were inconclusive. Confidence in all network comparisons was very low. Conclusions: Current randomized evidence does not establish a reliable clinical-effectiveness hierarchy among RB-ES, PM-ES, and CPM. Continuous monitoring showed directionally favorable but imprecise estimates for several outcomes. Surveillance strategies should therefore not be selected solely according to algorithm class or predictive complexity; their clinical effects may also depend on the target population, background monitoring, workflow integration, alert presentation, and clinical response pathway.
- Open Access
- 2 June 2026
- Published by Elsevier in SSRN Electronic Journal
CALMS, A Calibration-Optimised Deep Learning Model and Distilled Bedside Score for Mortality Prediction in Sepsis
Background: Existing prediction tools for sepsis mortality are limited by poor probability calibration, marked performance loss on external validation, and limited bedside applicability outside data-rich electronic health record environments. We aimed to develop and externally validate a calibration-optimised multimodal deep learning model for in- hospital mortality prediction in sepsis (CALMS) and to derive a parsimonious bedside score (qCALMS) by knowledge distillation, with validation in cohorts spanning two continents and three independent electronic health record systems.Methods: In this retrospective multicentre cohort study, we included adults (≥18 years) meeting the Sepsis-3 definition from three publicly available databases — MIMIC-IV v3.1 (development and internal validation; n=29 779), eICU-CRD v2.0 (external validation; n=21 558), and MIMIC-IV-ED v2.2 (emergency-department external validation of qCALMS; n=857) — and from an independent local cohort of 797 adult ICU admissions with Sepsis-3 from a tertiary teaching hospital in Changsha, China (Jan 1, 2019 to Dec 31, 2025). CALMS integrated static clinical characteristics, temporal vital sign and laboratory trajectories, Sequential Organ Failure Assessment (SOFA) subscores, and treatment interventions through parallel gated recurrent unit encoders with temporal attention, and was trained with a novel reversible mask-difference (RMD) augmentation strategy designed to mitigate cross-institutional distribution shift. The primary outcome was in-hospital mortality. Discrimination was assessed by area under the receiver operating characteristic curve (AUROC), calibration by expected calibration error (ECE) and Brier score, and clinical utility by decision-curve analysis. qCALMS was derived from CALMS-predicted probabilities through knowledge distillation, restricted to dichotomised laboratory variables routinely available within 6 h before ICU admission, and compared head-to-head with qSOFA and the systemic inflammatory response syndrome (SIRS) criteria.Findings: CALMS achieved an AUROC of 0·867 (95% CI 0·856–0·879) in the MIMIC-IV internal test set and 0·810 (0·802–0·817) in eICU-CRD, with the best probability calibration of all evaluated models in both cohorts (ECE 0·010 and 0·057; Brier scores 0·093 and 0·115). Cross-dataset performance degradation was 6·6% for CALMS versus 31·7% for an otherwise identical recurrent architecture trained without RMD augmentation, representing an approximately five-fold reduction in the external generalisation gap. Bayesian posterior analysis showed CALMS to be superior to all comparators (SOFA, qSOFA, logistic regression, XGBoost, GRU-Simple) with posterior probability ≥0·91 in both cohorts. The distilled qCALMS score, comprising blood urea nitrogen ≥30 mg/dL, bicarbonate <18 mEq/L, lactate ≥4 mmol/L, and platelet count <50 × 10⁹/L, stratified in-hospital mortality monotonically from 10–11% at score 0 to 55–58% at score 4 in both public cohorts (Cochran–Armitage p<0·0001), and significantly outperformed qSOFA and SIRS in head-to-head comparison (AUROC 0·645 [95% CI 0·623–0·667] in eICU-CRD and 0·619 [0·563–0·675] in MIMIC-IV-ED; both DeLong p<0·05). qCALMS computed from data available up to 6 h before ICU admission achieved discrimination comparable with SOFA computed 12 h after ICU admission (AUROC 0·626 vs 0·645), corresponding to an approximately 18-h early-warning lead time at equivalent discrimination. In the independent local cohort (n=797; 67 in-hospital deaths; event rate 8·4%), qCALMS retained robust discrimination (AUROC 0·635, 95% CI 0·567–0·702), excellent calibration (Brier 0·075; ECE 0·004; Hosmer–Lemeshow p=0·746), and significantly outperformed SIRS (AUROC 0·514, 95% CI 0·453–0·575; DeLong p=0·012), with mortality rising monotonically from 5·6% at score 0 to 23·5% at score 3.Interpretation: CALMS provides accurate and well-calibrated mortality prediction in sepsis with stable cross-institutional performance, and qCALMS translates this predictive capacity into a four-item laboratory-based bedside tool that outperforms qSOFA and SIRS, generalises to a fully independent non-US electronic health record system, and offers an approximate 18-h early-warning lead time over SOFA at equivalent discrimination. The dual-layer framework of developing a high-performance, calibration-optimised deep learning model and subsequently distilling its predictive knowledge into a parsimonious bedside score may be generalisable to other critical illness prediction tasks where the competing demands of model performance and bedside applicability must be reconciled. Prospective interventional studies are now required to determine whether qCALMS- guided early triage improves patient-centred outcomes.
- 2 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCALMS, A Calibration-Optimised Deep Learning Model and Distilled Bedside Score for Mortality Prediction in Sepsis
Background: Existing prediction tools for sepsis mortality are limited by poor probability calibration, marked performance loss on external validation, and limited bedside applicability outside data-rich electronic health record environments. We aimed to develop and externally validate a calibration-optimised multimodal deep learning model for in- hospital mortality prediction in sepsis (CALMS) and to derive a parsimonious bedside score (qCALMS) by knowledge distillation, with validation in cohorts spanning two continents and three independent electronic health record systems.Methods: In this retrospective multicentre cohort study, we included adults (≥18 years) meeting the Sepsis-3 definition from three publicly available databases — MIMIC-IV v3.1 (development and internal validation; n=29 779), eICU-CRD v2.0 (external validation; n=21 558), and MIMIC-IV-ED v2.2 (emergency-department external validation of qCALMS; n=857) — and from an independent local cohort of 797 adult ICU admissions with Sepsis-3 from a tertiary teaching hospital in Changsha, China (Jan 1, 2019 to Dec 31, 2025). CALMS integrated static clinical characteristics, temporal vital sign and laboratory trajectories, Sequential Organ Failure Assessment (SOFA) subscores, and treatment interventions through parallel gated recurrent unit encoders with temporal attention, and was trained with a novel reversible mask-difference (RMD) augmentation strategy designed to mitigate cross-institutional distribution shift. The primary outcome was in-hospital mortality. Discrimination was assessed by area under the receiver operating characteristic curve (AUROC), calibration by expected calibration error (ECE) and Brier score, and clinical utility by decision-curve analysis. qCALMS was derived from CALMS-predicted probabilities through knowledge distillation, restricted to dichotomised laboratory variables routinely available within 6 h before ICU admission, and compared head-to-head with qSOFA and the systemic inflammatory response syndrome (SIRS) criteria.Findings: CALMS achieved an AUROC of 0·867 (95% CI 0·856–0·879) in the MIMIC-IV internal test set and 0·810 (0·802–0·817) in eICU-CRD, with the best probability calibration of all evaluated models in both cohorts (ECE 0·010 and 0·057; Brier scores 0·093 and 0·115). Cross-dataset performance degradation was 6·6% for CALMS versus 31·7% for an otherwise identical recurrent architecture trained without RMD augmentation, representing an approximately five-fold reduction in the external generalisation gap. Bayesian posterior analysis showed CALMS to be superior to all comparators (SOFA, qSOFA, logistic regression, XGBoost, GRU-Simple) with posterior probability ≥0·91 in both cohorts. The distilled qCALMS score, comprising blood urea nitrogen ≥30 mg/dL, bicarbonate <18 mEq/L, lactate ≥4 mmol/L, and platelet count <50 × 10⁹/L, stratified in-hospital mortality monotonically from 10–11% at score 0 to 55–58% at score 4 in both public cohorts (Cochran–Armitage p<0·0001), and significantly outperformed qSOFA and SIRS in head-to-head comparison (AUROC 0·645 [95% CI 0·623–0·667] in eICU-CRD and 0·619 [0·563–0·675] in MIMIC-IV-ED; both DeLong p<0·05). qCALMS computed from data available up to 6 h before ICU admission achieved discrimination comparable with SOFA computed 12 h after ICU admission (AUROC 0·626 vs 0·645), corresponding to an approximately 18-h early-warning lead time at equivalent discrimination. In the independent local cohort (n=797; 67 in-hospital deaths; event rate 8·4%), qCALMS retained robust discrimination (AUROC 0·635, 95% CI 0·567–0·702), excellent calibration (Brier 0·075; ECE 0·004; Hosmer–Lemeshow p=0·746), and significantly outperformed SIRS (AUROC 0·514, 95% CI 0·453–0·575; DeLong p=0·012), with mortality rising monotonically from 5·6% at score 0 to 23·5% at score 3.Interpretation: CALMS provides accurate and well-calibrated mortality prediction in sepsis with stable cross-institutional performance, and qCALMS translates this predictive capacity into a four-item laboratory-based bedside tool that outperforms qSOFA and SIRS, generalises to a fully independent non-US electronic health record system, and offers an approximate 18-h early-warning lead time over SOFA at equivalent discrimination. The dual-layer framework of developing a high-performance, calibration-optimised deep learning model and subsequently distilling its predictive knowledge into a parsimonious bedside score may be generalisable to other critical illness prediction tasks where the competing demands of model performance and bedside applicability must be reconciled. Prospective interventional studies are now required to determine whether qCALMS- guided early triage improves patient-centred outcomes.
- 1 January 2021
- Published by Elsevier in Public Health
- Vol. 190, 82-88
Comparative efficacy of respiratory personal protective equipment against viral respiratory infectious diseases in healthcare workers: a network meta-analysis
- 1 January 2021
- Published by Elsevier in Public Health
- Vol. 190, 82-88
Comparative efficacy of respiratory personal protective equipment against viral respiratory infectious diseases in healthcare workers: a network meta-analysis
- Open Access
- 27 April 2020
- Published by SAGE Publications in Journal of International Medical Research
- Vol. 48 (4)
Outcomes and prognostic factors of tigecycline treatment for hospital-acquired pneumonia involving multidrug-resistant Acinetobacter baumannii
Objective: To compare outcomes and prognostic factors of tigecycline (TC)-based treatment with those of other antibiotic-based treatments in the treatment of hospital-acquired pneumonia caused by multidrug-resistant Acinetobacter baumannii (MDRAB). Methods: A retrospective analysis of data was performed from all patients ≥18 years who were treated in the ICU at Xiangya Hospital, Changsha, China (January 2016 to June 2017) with hospital-acquired pneumonia involving monomicrobial MDRAB. Patients were separated into TC and non-TC groups. Results: Of 86 MDRAB-positive patients, 59 were in the TC group and 27 were in the non-TC group. The 28-day death rates were not significantly different between the two groups, but the TC group had significantly more patients with a good clinical prognosis than the non-TC group. Although prognostic markers for a poor clinical response were sepsis, procalcitonin concentration and APACHE II scores, TC therapy was found to be a protective factor. Conclusions: TC based therapy was associated with a positive clinical response in the treatment of MDRAB caused hospital-acquired pneumonia. Further studies are required to confirm our results.
- 27 April 2020
- Published by SAGE Publications in Journal of International Medical Research
- Vol. 48 (4)
Open AccessOutcomes and prognostic factors of tigecycline treatment for hospital-acquired pneumonia involving multidrug-resistant Acinetobacter baumannii
Objective: To compare outcomes and prognostic factors of tigecycline (TC)-based treatment with those of other antibiotic-based treatments in the treatment of hospital-acquired pneumonia caused by multidrug-resistant Acinetobacter baumannii (MDRAB). Methods: A retrospective analysis of data was performed from all patients ≥18 years who were treated in the ICU at Xiangya Hospital, Changsha, China (January 2016 to June 2017) with hospital-acquired pneumonia involving monomicrobial MDRAB. Patients were separated into TC and non-TC groups. Results: Of 86 MDRAB-positive patients, 59 were in the TC group and 27 were in the non-TC group. The 28-day death rates were not significantly different between the two groups, but the TC group had significantly more patients with a good clinical prognosis than the non-TC group. Although prognostic markers for a poor clinical response were sepsis, procalcitonin concentration and APACHE II scores, TC therapy was found to be a protective factor. Conclusions: TC based therapy was associated with a positive clinical response in the treatment of MDRAB caused hospital-acquired pneumonia. Further studies are required to confirm our results.
- 6 March 2020
- Published by Wiley in Oral Diseases
- Vol. 26 (5) , 998-1009
H‐type blood vessels participate in alveolar bone remodeling during murine tooth extraction healing
Objectives: We aimed to investigate whether skeletal‐specific H‐type blood vessels exist in alveolar bone and how they function in alveolar bone remodeling.Materials and Methods: H‐type vessels with high expression of CD31 and Endomucin (CD31hiEmcnhi) were immunostained in alveolar bone. Abundance and age‐related changes in CD31hiEmcnhi endothelial cells (H‐ECs) were detected by flow cytometry. Osteoprogenitors association with H‐type vessels and bone mass were detected in tooth extraction model of alveolar bone remodeling by immunohistofluorescence and micro‐CT, respectively. Transcription and expression of H‐EC feature genes during in vitro Notch inhibition were measured by RT‐qPCR and immunocytofluorescence.Results: We verified that H‐type vessels existed in alveolar bone, the abundance of which was highest at infancy age, then decreased but maintained a constant level during aging. In tooth extraction model, H‐ECs significantly increased with concomitant perivascular accumulation of Runx2+ osteoprogenitors and gradually augmentation of bone mass. Notch inhibition of in vitro cultured H‐ECs resulted in decreased expression levels of Emcn and hes1, but not Pecam1 or Kdr genes, with decreased expression levels of H‐EC numbers, accordingly.Conclusions: The present study suggests that H‐type vessels promote osteogenesis during alveolar bone remodeling. Notch signaling pathway regulates expression of Emcn and possibly determines fate and functions of alveolar H‐ECs.
- 6 March 2020
- Published by Wiley in Oral Diseases
- Vol. 26 (5) , 998-1009
H‐type blood vessels participate in alveolar bone remodeling during murine tooth extraction healing
Objectives: We aimed to investigate whether skeletal‐specific H‐type blood vessels exist in alveolar bone and how they function in alveolar bone remodeling.Materials and Methods: H‐type vessels with high expression of CD31 and Endomucin (CD31hiEmcnhi) were immunostained in alveolar bone. Abundance and age‐related changes in CD31hiEmcnhi endothelial cells (H‐ECs) were detected by flow cytometry. Osteoprogenitors association with H‐type vessels and bone mass were detected in tooth extraction model of alveolar bone remodeling by immunohistofluorescence and micro‐CT, respectively. Transcription and expression of H‐EC feature genes during in vitro Notch inhibition were measured by RT‐qPCR and immunocytofluorescence.Results: We verified that H‐type vessels existed in alveolar bone, the abundance of which was highest at infancy age, then decreased but maintained a constant level during aging. In tooth extraction model, H‐ECs significantly increased with concomitant perivascular accumulation of Runx2+ osteoprogenitors and gradually augmentation of bone mass. Notch inhibition of in vitro cultured H‐ECs resulted in decreased expression levels of Emcn and hes1, but not Pecam1 or Kdr genes, with decreased expression levels of H‐EC numbers, accordingly.Conclusions: The present study suggests that H‐type vessels promote osteogenesis during alveolar bone remodeling. Notch signaling pathway regulates expression of Emcn and possibly determines fate and functions of alveolar H‐ECs.
- 1 August 2019
- Published by Elsevier in The American Journal of Emergency Medicine
- Vol. 37 (8) , 1601.e1-1601.e2
Transient visual impairment in a patient with psychogenic fever: A case report
- 1 August 2019
- Published by Elsevier in The American Journal of Emergency Medicine
- Vol. 37 (8) , 1601.e1-1601.e2
Transient visual impairment in a patient with psychogenic fever: A case report