Sage Journals: Discover world-class research

Abstract

Globus pallidus necrosis (GPN) is one of typical neurological imaging features in patients with carbon monoxide (CO) poisoning. Current clinical guideline recommends neurological imaging examination for CO-intoxicated patients with conscious disturbance rather than routine screening, which may lead to undiagnosed GPN. We aimed to develop an artificial intelligence algorithm for predicting GPN in CO intoxication patients. We included CO intoxication patients with neurological images between 2000 and 2019 in Chang Gung Memorial Hospital. We collected 41 clinical and laboratory parameters on the first day of admission for algorithm development. We used fivefold cross validation and applied several machine learning algorithms. Random forest classifier (RFC) provided the best predictive performance in our cohort. Among the 261 patients with CO intoxication, 52 patients presented with GPN. The artificial intelligence algorithm using the RFC-based AI model achieved an accuracy = 79.2 ± 2.6%, sensitivity = 77.7%, precision score = 81.9 ± 3.4%, and F1 score = 73.2 ± 1.8%. The area under receiver operating characteristic was approximately 0.64. Top five weighted variables were Platelet count, carboxyhemoglobin, Glasgow Coma scale, creatinine, and hemoglobin. Our RFC-based algorithm is the first to predict GPN in patients with CO intoxication and provides fair predictive ability. Further studies are needed to validate our findings.

Keywords

Artificial intelligence carbon monoxide globus pallidus necrosis random forest classifier

Introduction

Carbon monoxide (CO) poisoning is a major global public health problem. An estimation based on the Global Health Data Exchange registry showed 137 patients per million would suffer from CO intoxication and 4.6 CO-intoxicated patients per million would be fatal in 2017.¹ Such problem have become more severe since the beginning of the 21st century when the media widely reported charcoal burning as new painless suicide method.² In consequence, charcoal-burning suicide rate increased dramatically, especially in Hong Kong, Taiwan, Japan, Korea, and Singapore.³ In fact, the incidence of charcoal-burning suicide increased 29.5-fold in 2006 compared to in 1999 and 33.5% of suicide deaths in Taiwan were caused by this single suicide method.⁴ In addition to CO intoxication caused by charcoal-burning suicide, the incidence of accidental CO poisoning has also increased due to usage of indoor water heaters during cold weather.⁵ Therefore, CO intoxication is an issue that should not be ignored by clinicians.

Delayed neuropsychiatric syndrome (DNS) is one of the most debilitating sequelae after CO poisoning.⁶ Survivors may still experience cognitive impairment, akinetic mutism, sphincter incontinence, gait ataxia and extrapyramidal syndromes even after the acute stage.⁷ Imaging tools such as computer tomography (CT) and magnetic resonance imaging (MRI) have been developed to stratify high-risk CO-intoxicated patients that would develop DNS. Currently, the pathophysiology of CO-related DNS has not been fully elucidated. Several mechanisms, including impaired oxygen delivery, impaired oxygen utilization, oxidative stress with lipid peroxidation, and ischemia-reperfusion injury have been proposed for CO-related brain injury.^6,8–12 Hypoxic injury caused by impaired oxygen delivery and utilization are attributed to hemoglobin preferentially binding to CO instead of oxygen, prevention of oxygen release from hemoglobin into hypoxic environment, and shutdown of electron transportation chain due to CO binds to heme a3 subunit two of cytochrome C oxidase in mitochondria.^6,13,14 Notably, globus pallidus may be more susceptible to CO intoxication due to high iron content, high oxygen consumption, and poor anastomotic blood supply.^15,16 Globus pallidus necrosis (GPN) is a common brain lesion and associated with the development of DNS, which presents as symmetric hypodensity over globus pallidus in CT and low intensity on T1-weighted images but high intensity on T2-weighted images over medial portion of the globus pallidus on MRI.^17–20 Current guideline recommends brain CT or MRI for CO-intoxicated patients with loss of consciousness rather than routine screening.²¹ However, typical neurological imaging finding may still occur in asymptomatic CO-intoxicated patients.²² A point-of-care clinical prediction tool may help clinician efficiently sort out patients who are more likely to have GPN before performing the neuroimaging.

There has been tremendous interest regarding the utility of artificial intelligence (AI) in medical research. The utility of AI in the field of acute kidney injury (AKI) has been studied not only in bedside application, but also in improving our understanding of the disease.²³ Different AI algorithms have been utilized for different medical research with various degrees of accuracy, including toxicology studies.^24–27 AI prediction has the advantage of incorporating multiple parameters to improve predicative ability. Currently, there is no literature utilizing AI for predicting GPN in CO intoxication patients. We aim to determine the feasibility of using AI for predicting GPN in CO intoxication patients and determine which algorithm provides the best predictive ability.

Materials and methods

Ethics statement

This study complied with the guidelines of the Declaration of Helsinki and was approved by the Medical Ethics Committee of Chang Gung Memorial Hospital, Linkou, Taiwan. The institutional review board number was 201701109B0.

Patients

We included all patients with CO intoxication treated at Chang Gung Memorial Hospital between 2000 and 2019. Chang Gung Memorial Hospital is Taiwan's largest tertiary medical center, with 3700 beds and a specialized clinical poison center. The patients with available neurological imaging were included in the development and validation of artificial intelligence algorithms. We included a total of 261 patients in the cohort; the medical records of these patients, including demographic, clinical parameters, and laboratory data were integrated into the analysis.

Clinical management

The diagnosis of carbon CO intoxication was based on the exposure history and a blood carboxyhemoglobin level of more than 10% for smokers or 3% for non-smokers.²⁸ Supportive treatment and oxygen therapy with non-rebreathing facemask were provided for all CO intoxication patients. Due to hyperbaric oxygen (HBO) therapy is not covered in health reimbursement in Taiwan, the indication for HBO was not standardized in all patients. However, HBO therapy would not provide for patients with contraindications such as untreated pneumothorax, asthma, chronic obstructive pulmonary disease, seizure, pregnancy, high fever, upper respiratory tract infections, active malignancy, congenital spherocytosis, claustrophobia, and patients need intensive care.

Definitions of globus pallidus necrosis

The diagnosis of globus pallidus necrosis was based on typical radiographic imaging characteristics.¹⁸ The typical findings included symmetric hypodensity over globus pallidus in computer tomography, and low intensity on T1-weighted but high intensity on T2-weighted images over the bilateral globus pallidus area in MRI.

Statistical analysis

Data are expressed as the mean ± standard deviation or number (percentage), unless otherwise stated. t test was used to compare the means of continuous variables and normally distributed data. Categorical data was analyzed using the chi-square test. All statistical tests were two-tailed, with p values < .05 considered statistically significant. Data were analyzed with SPSS 25.0 software for Windows (SPSS, Inc., Chicago, IL, USA).

Data collection and feature engineering

The algorithm development process included three steps. First, demographics, vital signs, laboratory values, interventions, medications, and nurse documentation were accessed through the hospital information system of Chang Gung Memorial Hospital. The data collected is presented in the case report form (Supplementary Table S1). For missing values in the other recording data, we filled them with the median values of each corresponding feature. This process guarantees the quality and consistency of the data as well as the model. Next, we conducted standard normalization of every feature, alleviating the influence of a different range of values. We selected 41 clinical and laboratory features recorded on the first day of admission for use in algorithm development. Third, during feature engineering, the 261 patients were divided into training data (n = 209) and testing data (n = 52). The RFC-based AI model was developed using 80% of the records as the training cohort and 20% of the records as the testing cohort. All development and analyses associated with AI were performed using Jupyter Notebook Version 6.4.8.

Model selection

We applied several machine-learning methods, such as random forest classifier (RFC), K-nearest neighbor classification (KNN), logistic regression (LR), support vector machine, eXtreme gradient boosting (XGB), and neural network, and used fivefold cross-validation for the assessment of model performance. Cross-validation is a resampling procedure used to evaluate machine learning models, and “5” refers to the number of groups the given data sample was split into. We found that RFC, a decision-tree-based algorithm had the best performance. Therefore, we adopted RFC and LR for the following artificial intelligence algorithm development.

Performance evaluation

After we used random forest classifier and logistic regression to fit the training dataset and obtained the probability threshold, we computed the globus pallidus necrosis probability of the testing dataset as well as the accuracy, recall (sensitivity), specificity, precision and F1 score given that specific threshold. The F1 score provides an equal consideration to both precision and recall by calculating their harmonic mean into a single metric.²⁹ More importantly, we computed the receiver operating characteristic (ROC) curve for performance evaluation. We also performed a feature weight analysis to differentiate the contribution of each feature to the performance of the algorithm to evaluate the algorithm’s fitness for the clinical scenario. Feature weighting is closely associated with the AUROC, sensitivity, and precision of an algorithm and can be obtained during cross-validation in model training.

Results

Study population characteristics

Among the 261 patients with CO intoxication, 142 (54.4%) were male, and their mean age was 32.46 years (Table 1). 52 patients (19.9%) developed GPN. Baseline comorbidities such as hypertension, diabetes mellitus, coronary artery disease, respiratory disease, renal disease, and respiratory disease are similarly prevalent in both GPN patients and non-GPN patients. Patients with GPN not only had a higher incidence of fever episodes (44.2% vs 21.5%, p = .001) and respiratory failure (55.8% vs 28.7%, p < .001) but also experienced more musculoskeletal symptoms (40.4% vs 17.2%, p = .001) than patients without GPN at admission. Compared to the non-GPN group, both acute myocardial infarction (AMI) and AKI are more likely to occur in the GPN group (p = .018 and p = .02, respectively). The Glasgow Coma Scale is significantly lower in patients with GPN than those without (7.44 ± 4.56 vs 10.79 ± 4.78, p < .001). Although similar in baseline smoking habits, carboxyhemoglobin is significantly lower in the GPN patients than in patients without GPN (16.98 ± 17.18 vs 22.89 ± 16.84, p = .03). Serum sodium was higher in the GPN group than in the non-GPN group (144.52 ± 9.8 vs 141.35 ± 6.74, p = .031). A similar portion of patients were treated with hyperbaric oxygen therapy in GPN group and non-GPN group (21.2% vs 18.2%).

Table 1.

Clinical data of patients with carbon monoxide intoxication, stratified according to the status of globus pallidus necrosis (GPN) (n = 261).

Variable	Patients with GPN (n = 52)	Patients without GPN (n = 209)	p value
Baseline characteristics
Age, year	34.38 ± 16.76	31.98 ± 19.37	0.413
Male gender, n (%)	31 (59.6)	111 (53.1)	0.439
Smoking habit, n (%)	21 (40.4)	64 (30.6)	0.189
Hypertension, n (%)	1 (1.9)	14 (6.7)	0.371
Diabetes mellitus, n (%)	0 (0)	7 (3.3)	0.351
Coroanry artery disease, n (%)	1 (1.9)	10 (4.8)	0.699
Respiratory disease, n (%)	3 (5.8)	6 (2.9)	0.388
Renal disease, n (%)	2 (3.8)	6 (2.9)	0.661
Neurological disease, n (%)	3 (5.8)	5 (2.4)	0.199
Clinical manifestations
GCS	7.44 ± 4.56	10.79 ± 4.78	<0.001
Shock episode, n (%)	5 (9.6)	17 (8.1)	0.780
Fever episode, n (%)	23 (44.2)	45 (21.5)	0.001
Respiratory failure, n (%)	29 (55.8)	60 (28.7)	<0.001
AKI, n (%)	20 (38.5)	46 (22.0)	0.020
Neurological symptoms, n (%)	48 (92.3)	169 (80.9)	0.061
AMI, n (%)	9 (17.3)	12 (5.7)	0.018
Musculoskeletal symptoms, n (%)	21 (40.4)	36 (17.2)	0.001
Laboratory findings
COHb (%)	16.98 ± 17.18	22.89 ± 16.84	0.03
pH	7.35 ± 0.13	7.36 ± 0.13	0.736
PCO₂, mmHg	35.60 ± 8.09	35.73 ± 10.00	0.928
PO_2, mmHg	180.92 ± 145.63	160.70 ± 124.80	0.313
HCO₃, mmol/L	19.54 ± 4.52	20.03 ± 4.86	0.515
BE, mEq/L	−5.84 ± 6.33	−4.41 ± 6.03	0.133
RBC	4.64 ± 0.68	4.59 ± 0.73	0.646
Hemoglobin, g/dL	13.91 ± 2.51	13.81 ± 2.54	0.798
WBC, 10³/μL	23.16 ± 9.86	25.08 ± 9.50	0.193
Platelet, 10³/μL	449.38 ± 2093.81	149.25 ± 69.73	0.306
Segment count, %	78.63 ± 15.69	78.53 ± 21.88	0.975
Creatinine, mg/dL	1.90 ± 2.46	1.44 ± 1.88	0.142
Sodium, mEq/L	144.52 ± 9.80	141.35 ± 6.74	0.031
Potassium, mEq/L	4.08 ± 0.69	4.06 ± 0.98	0.894
ALT > 2* upper limit, n (%)	17 (32.7)	31 (14.8)	0.005
Treatment and outcome
HBO, n (%)	11 (21.2)	38 (18.2)	0.692
Mortality, n (%)	4 (7.7)	15 (7.2)	1.000

AKI: acute kidney injury; ALT: alanine aminotransferase; AMI: acute myocardial injury; AST: aspartate aminotransferase; BE: base excess; COHb: carboxyhemoglobin; GCS: glascow coma scale; HBO: hyperbaric oxygen; HCO₃: bicarbonate; PCO₂: partial pressure of carbon dioxide; PO₂: partial pressure of oxygen; WBC: white blood cell.

The performance of developed AI algorithms

The established algorithm achieved the following results: accuracy = 79.2 ± 2.6%, sensitivity = 77.7%, precision = 81.9 ± 3.4%, and F1 = 73.2 ± 1.8%. The AUROC was approximately 0.64. The weight features were calculated, and the top 10 were listed according to their categories (Table 2). The weight represents the influence of each feature. All clinical features and their importance are illustrated in the Figure 1. Platelet count, carboxyhemoglobin, Glasgow Coma scale, creatinine, and hemoglobin were the most highly weighted.

Table 2.

Identification of top ten weighted features for globus pallidus necrosis (GPN) using artificial intelligence approach.

Feature variable	Weight
Platelet	0.07244399
COHb	0.06966668
GCS	0.06808823
Creatinine	0.05911923
Hemoglobin	0.05359761
Sodium	0.05258619
PO2	0.04987526
Segment	0.04935926
RBC	0.04845507
WBC	0.04547123

COHb: carboxyhemoglobin; GCS: glascow coma scale; PO₂: partial pressure of oxygen; RBC: red blood cell; WBC: white blood cell.

Figure 1.

Ranking the weight of each variable in the random forest classifier algorithm. AMI: acute myocardial injury; ARF: acute renal failure; BE: base excess; COHb: carboxyhemoglobin; CVSym: cardiovascular symptoms; DM: diabetes mellitus; GCS: glascow coma scale; GIsym: gastrointestinal symptoms; HBO: hyperbaric oxygen; HCO₃: bicarbonate; hepatitisALT2X: patients with alanine aminotransferase increase more than 2 times of upper limit; ICU: intensive care unit admission; K: potassium; MusculoSym: musculoskeletal symptoms; neurologicalD: baseline neurological comorbidities; NeuroSym: neurological symptoms; PCO₂: partial pressure of carbon dioxide; PH: blood pH value; PO₂: partial pressure of oxygen; RBC: red blood cell; renalD: baseline renal comorbidities; ResFailure: respiratory failure; respiratoryD: baseline respiratory comorbidities; WBC: white blood cell bicarbonate; PCO₂: partial pressure of carbon dioxide; PO₂: partial pressure of oxygen; WBC: white blood cell.

Discussion

To our best knowledge, this is the first study to predict globus pallidus necrosis after carbon monoxide intoxication using RFC-based artificial intelligence. Our results demonstrated that the established algorithm can achieve an accuracy of 79.2% and precision of 81.9%. The top five weighted variables for predicting GPN in our algorithm are platelet count, carboxyhemoglobin, Glasgow Coma scale, creatinine, and hemoglobin (Figure 1 and Table 2).

The development of neurological sequela is a long-term burden for survivors of CO intoxication. Ku et al. reported 59.7% of patients developed GPN after CO intoxication in Taiwan, but the incidence rate ranging from 18.7% to 100% has been reported by different studies.^20,30,31 Current guideline from Centers for Disease Control and Prevention recommends brain imaging workup for CO-intoxicated patients with loss of consciousness or cardiopulmonary signs and symptoms.²¹ However, typical neurological imaging findings may still occur in asymptomatic CO-intoxicated patients.²² Considering DNS is associated with GPN, identification of those who are at higher risk of developing GPN is crucial for CO intoxication patients. Although several predictors have been investigated by previous group,^30,32 no AI-based computerized tools have been developed to predict GPN until this study. Our algorithm provides an adequate prediction ability to predict GPN with clinical characteristics that can be easily acquired by clinicians and therefore the implementation into clinical practice would be easy. The features used in this AI algorithm are clinical or laboratory data obtained during the routine management of CO intoxication patients, without imposing any additional burden on frontline healthcare providers in the emergency department. Furthermore, by inputting these variables, the AI algorithm can provide a recommendation to clinicians on whether a neuroimaging examination is necessary within minutes, thereby alleviating the clinical workload for busy healthcare providers. Additionally, this algorithm has also opened up possibilities for future AI-related research in toxicology. It has confirmed the feasibility of using clinical features to predict toxicological outcomes through AI, paving the way for further studies in this area. Future studies are required to validate our results.

Platelet count, carboxyhemoglobin, GCS, creatinine, and hemoglobin are the top five weighted variables in the RFC model for GPN prediction. Despite our incomplete understanding of why these variables are weighted in this particular order in our algorithm, they have all been either reported to be predictors for GPN or pathophysiological related.^30,32,33 For example, creatinine has been reported as a predictor for GPN in charcoal-burning CO intoxication patients during univariate analysis.³⁰ In addition, higher platelet count has been reported in patients with CO intoxication compared to those without.³³ CO displaces NO from platelet surface hemoproteins and the displaced NO subsequently forms peroxynitrite, which leads to platelet activation.^34,35 Increased thrombotic events such as deep vein thrombosis have been reported in patients with CO intoxication.^36–38 Platelet activation also elicits an inflammatory effect via neutrophil activation. The level of cerebrospinal fluid myelin basic protein, a product of lipid peroxidation after neutrophil activation, is significantly higher in CO-intoxicated patients with DNS compared to those without.^39,40 Moreover, a meta-analysis including 2328 patients from 10 studies demonstrated that patients with initial low GCS score had higher risk for developing DNS compared to those with higher GCS score.⁴¹ Future studies may explore the different contributions of each variable to GPN.

CO exhibits an affinity for hemoglobin that is 200 times higher than that of oxygen, resulting in the formation of carboxyhemoglobin and causing a leftward shift in the oxyhemoglobin dissociation curve.⁴² Carboxyhemoglobin level is important for the diagnosis of CO intoxication but the level can be affected by the alveolar ventilation, blood volume, metabolic activity, smoking and duration of exposure.⁴² While a carboxyhemoglobin level above 50% is generally considered fatal, it does not reliably indicate the severity of systemic effects, poor prognosis, or the development of DNS.^42,43 The disparity between carboxyhemoglobin levels and the extent of CO poisoning could be attributed to variations in the time interval between patients' exposure to CO and their arrival at the hospital.⁴⁴ In other word, patients who experience lower carboxyhemoglobin levels but develop GPN might have experienced a longer delay in reaching the hospital and consequently had a lengthier exposure to ambient oxygen. However, obtaining the exact time of CO intoxication is challenging, if not impossible, as is the case in all real-world studies of CO poisoning.⁴⁵ Therefore, prediction of development GPN solely based on carboxyhemoglobin level is unreliable and our algorithm based on multiple clinical features can more effectively predict the occurrence of GPN.

Although not weighted in the top five, other previous reported predictors including acute myocardial injury, serum bicarbonate concentration, and age are also moderately weighted in our model. Patient’s underlying comorbidities, such as neurological disease, respiratory disease and renal disease, are also weighted in our model. The advantage of using AI is the ability to take all these variables into one prediction algorithm to improve predictive ability.

Our method of using an AI algorithm in analyzing intoxication patients has several advantages compared to using traditional statistics. First, all variables used in our algorithm are clinically relevant data and would be rapidly acquired while CO-intoxicated patients presented to the emergency department. In addition, the established algorithm is ready for clinical utility and swiftly predicts the risk of GPN. Implementation into clinical practice is easy. Moreover, the algorithm uses feature weighting to approximate the influence of individual features during AI classifier development, there is no need to manually discard variables as in the traditional statistical methods. Finally, the predictive power using AI is based on multiple variables rather than single variable as in traditional statistical method, which may provide better predictive ability. Although promising, our study is limited by the small sample size. Another limitation is that our study cannot extrapolate the relationship or causality among all clinical and laboratory features. Further studies are needed to validate our findings.

Conclusion

Our AI algorithm using the RFC method demonstrates fair ability to predict GPN in patients with CO intoxication. To our knowledge, this is the first AI-based prediction tool for patients with CO intoxication. The top five weighted features for GPN are Platelet count, carboxyhemoglobin, GCS, creatinine, and hemoglobin.

Supplemental Material

Supplemental Material - An artificial intelligence algorithm for analyzing globus pallidus necrosis after carbon monoxide intoxication

Supplemental Material for An artificial intelligence algorithm for analyzing globus pallidus necrosis after carbon monoxide intoxication by Ming-Jen Chan, Ching-Chih Hu, Wen-Hung Huang, Ching-Wei Hsu, Tzung-Hai Yen, and Cheng-Hao Weng in Human & Experimental Toxicology.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported by grants from Chang Gung Memorial Hospital Research Program (CMRPG5L0201 and CMRPG3L0991).

ORCID iDs

Ming-Jen Chan

Tzung-Hai Yen

Cheng-Hao Weng

Supplemental Material

Supplemental material for this article is available online.

References

Mattiuzzi

Lippi

. Worldwide epidemiology of carbon monoxide poisoning. Hum Exp Toxicol 2020; 39: 387–392. DOI: 10.1177/0960327119891214.

Chen

Tsai

Biddle

, et al. Newspaper reporting and the emergence of charcoal burning suicide in Taiwan: a mixed methods approach. J Affect Disord 2016; 193: 355–361. DOI: 10.1016/j.jad.2015.12.041 (accessed 06 January 2016).

Chang

Chen

Yip

, et al. Regional changes in charcoal-burning suicide rates in East/Southeast Asia from 1995 to 2011: a time trend analysis. PLoS Med 2014; 11: e1001622. DOI: 10.1371/journal.pmed.1001622 (accessed 01 April 2014).

Pan

Y-J

Liao

S-C

Lee

M-B

. Suicide by charcoal burning in Taiwan, 1995–2006. J Affect Disord 2010; 120: 254–257. DOI: 10.1016/j.jad.2009.04.003.

Shie

. Population-based case-control study of risk factors for unintentional mortality from carbon monoxide poisoning in Taiwan. Inhal Toxicol 2007; 19: 905–912. DOI: 10.1080/08958370701432173.

Rose

Wang

, et al. Carbon monoxide poisoning: pathogenesis, management, and future directions of therapy. Am J Respir Crit Care Med 2017; 195: 596–606. DOI: 10.1164/rccm.201606-1275CI.

Hsiao

Kuo

Huang

. Delayed encephalopathy after carbon monoxide intoxication--long-term prognosis and correlation of clinical manifestations and neuroimages. Acta Neurol Taiwan 2004; 13: 64–70.

Hardy

Thom

. Pathophysiology and treatment of carbon monoxide poisoning. J Toxicol Clin Toxicol 1994; 32: 613–629. DOI: 10.3109/15563659409017973.

Tritapepe

Macchiarelli

Rocco

, et al. Functional and ultrastructural evidence of myocardial stunning after acute carbon monoxide poisoning. Crit Care Med 1998; 26: 797–801. DOI: 10.1097/00003246-199804000-00034.

10.

Thom

. Carbon monoxide-mediated brain lipid peroxidation in the rat. J Appl Physiol 1990; 68: 997–1003. DOI: 10.1152/jappl.1990.68.3.997.

11.

Thom

Bhopale

Fisher

, et al. Delayed neuropathology after carbon monoxide poisoning is immune-mediated. Proc Natl Acad Sci USA 2004; 101: 13660–13665. DOI: 10.1073/pnas.0405642101 (accessed 01 October 2004).

12.

Weaver

. Carbon monoxide poisoning. Crit Care Clin 1999; 15: 297–317. DOI: 10.1016/s0749-0704(05)70056-7.

13.

Gnaiger

Lassnig

Kuznetsov

, et al. Mitochondrial oxygen affinity, respiratory flux control and excess capacity of cytochrome c oxidase. J Exp Biol 1998; 201: 1129–1139. DOI: 10.1242/jeb.201.8.1129.

14.

Lo Iacono

Boczkowski

Zini

, et al. A carbon monoxide-releasing molecule (CORM-3) uncouples mitochondrial respiration and modulates the production of reactive oxygen species. Free Radic Biol Med 2011; 50: 1556–1564. DOI: 10.1016/j.freeradbiomed.2011.02.033 (accessed 05 March 2011).

15.

Beppu

. The role of MR imaging in assessment of brain damage from carbon monoxide poisoning: a review of the literature. AJNR Am J Neuroradiol 2014; 35: 625–631. DOI: 10.3174/ajnr.A3489 (accessed 18 April 2013).

16.

Hurley

Hopkins

Bigler

, et al. Applications of functional imaging to carbon monoxide poisoning. J Neuropsychiatry Clin Neurosci 2001; 13: 157–160. DOI: 10.1176/jnp.13.2.157.

17.

O’Donnell

Buxton

Pitkin

, et al. The magnetic resonance imaging appearances of the brain in acute carbon monoxide poisoning. Clin Radiol 2000; 55: 273–280. DOI: 10.1053/crad.1999.0369.

18.

Chen

Lee

, et al. Brain injury after acute carbon monoxide poisoning: early and late complications. AJR Am J Roentgenol 2007; 189: W205–211. DOI: 10.2214/ajr.07.2425.

19.

Chu

Jung

Kim

, et al. Diffusion-weighted MRI and 99mTc-HMPAO SPECT in delayed relapsing type of carbon monoxide poisoning: evidence of delayed cytotoxic edema. Eur Neurol 2004; 51: 98–103. DOI: 10.1159/000076536 (accessed 28 January 2004).

20.

Moon

Chun

Baek

, et al. Initial diffusion-weighted MRI and long-term neurologic outcomes in charcoal-burning carbon monoxide poisoning. Clin Toxicol (Phila) 2018; 56: 161–169. DOI: 10.1080/15563650.2017.1352098 (accessed 28 July 2017).

21.

Prevention

CfDca.

Clinical guidance for carbon monoxide (CO) poisoning. Atlanta, GA: Centers for Disease Control and Prevention, 2022. Available at https://www.cdc.gov/disasters/co_guidance.html

22.

Durak

Coskun

Yikilmaz

, et al. Magnetic resonance imaging findings in chronic carbon monoxide intoxication. Acta Radiol 2005; 46: 322–327. DOI: 10.1080/02841850510021085.

23.

Gottlieb

Samuel

Bonventre

, et al. Machine learning for acute kidney injury prediction in the intensive care unit. Adv Chronic Kidney Dis 2022; 29: 431–438. DOI: 10.1053/j.ackd.2022.06.005.

24.

Bajaj

Koyner

. Artificial intelligence in acute kidney injury prediction. Adv Chronic Kidney Dis 2022; 29: 450–460. DOI: 10.1053/j.ackd.2022.07.009.

25.

Tangri

Ferguson

. Artificial intelligence in the identification, management, and follow-up of CKD. Kidney360 2022; 3: 554–556. DOI: 10.34067/kid.0007572021.

26.

Yen

Huang

, et al. An artificial intelligence algorithm for analyzing acetaminophen-associated toxic hepatitis. Hum Exp Toxicol 2021; 40: 1947–1954. DOI: 10.1177/09603271211014587 (accessed 06 May 2021).

27.

Oyaga-Iriarte

Insausti

Sayar

, et al. Prediction of irinotecan toxicity in metastatic colorectal cancer patients based on machine learning models with pharmacokinetic parameters. J Pharmacol Sci 2019; 140: 20–25. DOI: 10.1016/j.jphs.2019.03.004 (accessed 04 May 2019).

28.

Ernst

Zibrak

. Carbon monoxide poisoning. N Engl J Med 1998; 339: 1603–1608. DOI: 10.1056/NEJM199811263392206.

29.

Taha

Hanbury

. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC Med Imaging 2015; 15: 29. DOI: 10.1186/s12880-015-0068-x.

30.

Huang

Hsu

, et al. Incidence rate and predictors of globus pallidus necrosis after charcoal burning suicide. Int J Environ Res Public Health 2019; 16: 4426. DOI: 10.3390/ijerph16224426 (accessed 12 November 2019).

31.

Kim

Lee

Park

, et al. Acute carbon monoxide poisoning: MR imaging findings with clinical correlation. Diagn Interv Imaging 2017; 98: 299–306. DOI: 10.1016/j.diii.2016.10.004 (accessed 24 November 2016).

32.

Kim

Durey

Han

, et al. Predictive factors for acute brain lesions on magnetic resonance imaging in acute carbon monoxide poisoning. Am J Emerg Med 2020; 38: 1825–1830. DOI: 10.1016/j.ajem.2020.05.076 (accessed 28 May 2020).

33.

Karabacak

Varol

Türkdogan

, et al. Mean platelet volume in patients with carbon monoxide poisoning. Angiology 2014; 65: 252–256. DOI: 10.1177/0003319713497422 (accessed 30 July 2013).

34.

Thom

Ohnishi

Ischiropoulos

. Nitric oxide released by platelets inhibits neutrophil B2 integrin function following acute carbon monoxide poisoning. Toxicol Appl Pharmacol 1994; 128: 105–110. DOI: 10.1006/taap.1994.1186.

35.

Thom

Ischiropoulos

. Vascular endothelial cells generate peroxynitrite in response to carbon monoxide exposure. Chem Res Toxicol 1997; 10: 1023–1031. DOI: 10.1021/tx970041h.

36.

Chung

Lin

Kao

. Carbon monoxide poisoning and risk of deep vein thrombosis and pulmonary embolism: a nationwide retrospective cohort study. J Epidemiol Community Health 2015; 69: 557–562. DOI: 10.1136/jech-2014-205047 (accessed 22 January 2015).

37.

Lee

Lim

Kang

, et al. Pulmonary thromboembolism after carbon monoxide poisoning. Am J Emerg Med 2018; 36: e1713–e1717. DOI: 10.1016/j.ajem.2018.05.063 (accessed 29 May 2018).

38.

De Gregorio

Olivotto

Migliorini

, et al. Life-threatening acute pulmonary thromboembolism following severe carbon monoxide poisoning: a plausible association. JACC: Case Reports 2019; 1: 208–212. DOI: 10.1016/j.jaccas.2019.05.024.

39.

Thom

Bhopale

Han

, et al. Intravascular neutrophil activation due to carbon monoxide poisoning. Am J Respir Crit Care Med 2006; 174: 1239–1248. DOI: 10.1164/rccm.200604-557OC (accessed 24 August 2006).

40.

Kuroda

Fujihara

Kushimoto

, et al. Novel clinical grading of delayed neurologic sequelae after carbon monoxide poisoning and factors associated with outcome. Neurotoxicology 2015; 48: 35–43. DOI: 10.1016/j.neuro.2015.03.002 (accessed 07 March 2015).

41.

Namgung

Ahn

, et al. Association between glasgow coma scale in early carbon monoxide poisoning and development of delayed neurological sequelae: a meta-analysis. J Pers Med 2022; 12: 635. DOI: 10.3390/jpm12040635 (accessed 14 April 2022).

42.

Weaver

. Clinical practice. Carbon monoxide poisoning. N Engl J Med 2009; 360: 1217–1225. DOI: 10.1056/NEJMcp0808891.

43.

Palmeri

Gupta

. Carboxyhemoglobin toxicity. Treasure Island, FL: StatPearls Publishing, 2022.

44.

Manaker

Perry

. Carbon monoxide poisoning, Waltham, MA: UpToDate, 2023 (Accessed on 14 June 2023).

45.

Balzan

Agius

Galea Debono

. Carbon monoxide poisoning: easy to treat but difficult to recognise. Postgrad Med J 1996; 72: 470–473. DOI: 10.1136/pgmj.72.850.470.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.11 MB