Using machine learning methods to predict the lactate trend of sepsis patients in the ICU

doi:10.21203/rs.3.rs-1855422/v2

Download PDF

Article

Using machine learning methods to predict the lactate trend of sepsis patients in the ICU

https://doi.org/10.21203/rs.3.rs-1855422/v2

This work is licensed under a CC BY 4.0 License

Version 2

posted

You are reading this latest preprint version

Purpose

Serum lactate levels are considered a biomarker of tissue hypoxia. In sepsis or septic shock patients as suggested by The Surviving Sepsis Campaign serum lactate levels should be assessed and early lactate clearance-directed therapy is associated with decreased mortality. Monitoring a patient's vital parameters and repeatedly done blood analysis may have deleterious effects on the patient and brings an economical burden. Machine learning algorithms and trend analysis are gaining importance to overcome these unwanted facts. In this context, we aimed to investigate if an artificial intelligence approach can predict lactate trends from non-invasive clinical variables of patients with sepsis.

Methods

In this retrospective study, adult patients with sepsis from the MIMIC-IV dataset who had at least two serum lactate measurements recorded within the first 6 hours of sepsis diagnosis and who also has an ICU length of stay ≥ 24 hours are evaluated and ≥1mmol/l change is considered as a trend indicator. For prediction of lactate trend Naïve Bayes, J48 Decision Tree, Logistic Regression, Random Forest, and Logistic Model Tree (LMT) classifiers are evaluated.

Results

LMT algorithm outperformed other classifiers (AUC= 0.832). J48 decision tree performed worse when predicting constant lactate trend. LMT algorithm with 4 features (heart rate, oxygen saturation, lactate value before sepsis diagnosis, and time interval variables) achieved 0.821 in terms of AUC.

Conclusion

We can say that machine learning models that employ logistic regression architectures, i.e. LMT algorithm achieved good results in lactate trend prediction tasks can be effectively used to assess the state of the patient whether it is stable or improving.

Sepsis

serum lactate value

machine learning

intensive care unit

Serum lactate level has been traditionally considered a biomarker of tissue hypoxia and is often accompanied by sepsis [1]. Measuring and monitoring blood lactate concentration in sepsis and septic shock can reflect the severity of the illness and the response to therapeutic interventions [2–4]. It has been shown that the decrease in blood lactate values measured in the first hours of admission to the intensive care unit (ICU) over time is associated with better survival [5]. Persistently high or increased lactate levels as a marker of hypoperfusion are associated with increased morbidity and mortality [6].

For adults with sepsis or septic shock, The Surviving Sepsis Campaign (SSC) international consensus guideline suggests guiding resuscitation to decrease serum lactate in patients with elevated lactate levels over not using serum lactate [7]. Recently, some randomized control trials demonstrated that early lactate clearance-directed therapy is associated with decreased mortality as compared with the usual care [8]. Because the lactate level measurement is based on laboratory analysis, technologies that can predict lactate trends quickly, accurately, and noninvasively can be of great help to clinicians. Despite extensive efforts over the years, there are currently no commercially available IV chemical sensors (i.e., in the bloodstream) for continuous real-time monitoring of lactate levels in ICU patients [9]. Serum lactate measurements made at frequent intervals pose a risk of infection due to the need for venipuncture or the use of a central venous catheter for blood sampling [10], and risk of anemia due to repeated blood samplings [11]. In low- and middle-income countries (LMICs), continuous serum lactate measurements may not be available in intensive care units, or in the presence of limited resources, which may increase the financial burden [12]. A non-invasive method that could predict the course of serum lactate values without the need for repeated blood draws could alert clinicians to potential deterioration and perform rapid confirmatory testing by a blood draw. In addition, it may avoid unnecessary blood sampling and repeated lactate measurements. Machine learning algorithms may be helpful to clinicians in this regard [13].

We performed this retrospective study with the hypothesis that an artificial intelligence approach using machine learning can predict lactate trends from non-invasive clinical variables of patients with sepsis.

2.1. Data sources

The Medical Information Mart for Intensive Care (MIMIC- IV, v1.0) is a longitudinal, single-center database maintained by the Laboratory for Computational Physiology at the Massachusetts Institute of Technology (MIT) which contains de-identified data for over 60,000 patients admitted to ICUs at a US tertiary academic medical center (the Beth Israel Deaconess Medical Center [BIDMC]) [14]. We received permission to use the database, and all reporting followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines [15]. The datasets generated and/or analysed during the current study are available in the MIMIC Code Repository, (https://mimic-iv.mit.edu). STROBE guideline focuses on reporting results of observational studies without clear discussions of reporting ML analyses. For this reason, the recommendations for reporting and machine learning analyses in clinical research presented by Stevens et al.[16] were used as a template when preparing and submitting our manuscript. In this research statistical methods to be used in machine learning analysis in clinical research are recommended and machine learning analysis workflow is overviewed. Also, several key reporting elements according to the study designs are reported. The project is approved by the institutional review boards of MIT (Cambridge, MA, USA) and BIDMC (Boston, MA, USA) and was granted a waiver of informed consent. One of our authors (M.K.A., Record ID: 35109067) gained permission to document the database after online training at The Collaborative Institutional Training Initiative (CITI Program). Our research was conducted entirely on publicly available, anonymized data; therefore, individual patient consents were not required. All methods were carried out in accordance with relevant guidelines and regulations in the declarations to protect the privacy of patients.

Data availability: The data that support the findings of this study are available from MIMIC-IV but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Due to the data use agreement for the MIMIC-IV (v1.0) (PhysioNet Credentialed Health Data Use Agreement 1.5.0) there are restrictions that apply to the availability of these data, which were used under license for the current study, and so are not publicly available and we cannot share access to PhysioNet restricted data with anyone. (https://physionet.org/content/mimiciv/1.0/)

2.2 Study design

We conducted this retrospective study based on a subgroup of adult patients from the MIMIC IV dataset with the Sepsis-3 criteria; sepsis was defined as a suspected infection combined with an acute increase in Sequential/Sepsis-related Organ Failure Assessment (SOFA) score ≥ 2 [17].

We retrospectively evaluated a subgroup of adult patients (age ≥ 18 years) with sepsis from the MIMIC-IV dataset who had at least two serum lactate measurements recorded within the first 6 hours of sepsis diagnosis and who also has an ICU length of stay ≥ 24 hours.

2.3 Deﬁnition of outcomes

To accomplish this aforementioned lactate trend analysis, we first need to define trends. Three trend states are constructed according to value change in blood lactate. For the 24-hour observation period, 1mmol per liter and above change is considered a trend indicator. We calculated the change between two lactate values measured with a maximum interval of 6 hours. According to this setup, all samples in the data cohort have been labeled as increase, decrease, or constant. Trend definition can be seen in Fig.1.

2.5 Variable selection

We identified 9 variables that are most relevant in lactate trend analysis according to the clinical literature. These variables are age, first lactate value, last lactate value and the time interval between two lactate measurements, the mean of hemodynamic and respiratory monitoring parameters measured in this time interval (heart rate, systolic blood pressure, diastolic blood pressure, mean blood pressure, oxygen saturation, and PaO₂/FiO₂ ratio) (Table 1).

The main reason behind these variable selections is to reduce laboratory dependence on lactate trend analysis and do this analysis in a noninvasive manner. Preprocessing is a vital step to achieving robust machine learning models. These processes help reduce noise, remove redundant data, generate consistent data, and thus increase the performance of prediction models. To increase data quality, we applied some preprocessing steps to the data cohort. Outliers in the dataset are removed to obtain consistency between data points. To make the range intervals more coherent unity-based normalization is applied. All ranges are transformed into 0 and 1. We obtained 18653 data samples after these preprocessing steps.

Feature selection strategies on clinical data provide the right parameters to analyze a certain disease, treatment cost reduction, and reduce computational burden [18]. To achieve these goals, we do a further investigation on variable space. We used the Correlation-based feature selection (CFS) algorithm as a feature selector. CFS algorithm acquires important and pertinent features using inner characteristics of data instead of using machine learning approaches [19]. In many cases, some features have a high correlation with others. These features with high correlation characteristics produce redundant data and thus reduce the performance of prediction models. CFS algorithm evaluates the correlations between other features and discards features that have high correlation [19]. According to the CFS algorithm, we identified four variables that have less correlation than the others and can be used to predict lactate trends in sepsis patients. These are heart rate, oxygen saturation, lactate value before sepsis diagnosis, and time interval. The overall ranking of all features can be seen in Fig. 2. A pyramid shape is used to demonstrate the rank of features. In this pyramid, high-rank features appear on the top side whereas low-rank features appear on the bottom side of the pyramid.

2.5 Proposed Machine Learning Framework

Our proposed machine learning-based framework uses a clinical and demographical type of data and feeds these data to a classifier to oversee lactate trend in ICU settings. We utilized a traditional model for a supervised classification problem that consists of two parts: a training and a test/evaluation phase.

First of all, training data that consist of annotated data samples are acquired from the MIMIC IV dataset. Afterward, they go through a data preprocessing stage to increase data quality for the classification model. Every sample in training data has a lactate trend label (Increase/Constant/Decrease). These samples are trained with a classifier to construct a machine learning model. For the test stage, a preprocessed test sample is fed to the classifier and the classifier predicts its lactate trend label. In conclusion, classification performance is reported in the evaluation phase (Fig.3).

2.6 Selected Classifiers for Proposed Framework

For the prediction of lactate trends in sepsis patients, we evaluated several classifiers on the MIMIC IV dataset. These classifiers are Naïve Bayes (NB), J48 Decision Tree, Logistic Regression (LR), Random Forest (RF), and Logistic Model Tree (LMT). Naïve Bayes is a traditional and simple machine learning approach that contemplates dataset attributes as an independent [18]. The outputs are considered as class probabilities. Naïve Bayes acts on the Bayes theorem, which is the probability of any particular event occurring, given the probability of another event just occurring. The class with the highest probability is selected as the outcome. It became very popular in the machine learning area due to its advantages. These advantages are handling the overfitting problem very well and parallelization of the classification process [20]. J48 decision tree algorithm is an updated version of the popular decision tree algorithm ID3[21]. It can be used in both numerical and categorical data. J48 aims to find a specific attribute that fully partitions the training data. This attribute has the highest information gain value in the dataset [22]. By evaluating the probable values of this attribute, a branch pruning process starts and J48 defines target values. In the meantime, J48 searches other high information gain attributes. This process continues until an explicit decision is made on the combination of attributes that gives a certain rule for determining the target value. At the end of the algorithm, all attributes are evaluated and therefore all samples have a target value accordingly [22]. J48 became a popular machine learning tool in many areas due to its easy implementable and robust nature [21, 23, 24].

Random Forest (RF) belongs to the family of decision trees that employ a supervised ensemble learning strategy [25]. It gained popularity among classification and regression problem domains due to its robustness against overfitting and low computational load [25-27]. RF builds many decision trees that are based on the selection of a random subset of variables that are called bootstrap samples. Other decision tree learners aim to find the best variable available whereas RF uses random variables. The main motivation for this approach is to reduce the correlation between these candidate random trees. This randomness approach is important when making decisions because if there are highly correlated variables available then it affects the prediction phase and leads to poor prediction performance. As an output, all predictions from random trees are combined to achieve the utmost result [26].

The logistic regression algorithm is mainly used for tackling classification problems and modeling class probabilities [28]. It aims to fit the data to a logistic curve to predict the occurrence probability of events [29]. It can handle nonlinear dataset effects.

LMT algorithm is a hybrid decision tree approach that utilizes logistic regression and decision tree learning [30]. Leaves of the tree have piecewise linear regression models that are constructed by logistic regression functions. To construct these logistic regression functions LogitBoost algorithm is used [31]. Pruning of the decision tree is done by decision tree classifier algorithms. Splitting of the decision tree is implemented via logistic variant information gain. The algorithm has many positive aspects, it can map linear relationships, overfitting can be easily avoided, and it is easy to implement. Because of its numerous advantages, in recent years it has been used in many different research areas [30-32]

Ethics approval and consent to participate

Our research was conducted entirely on publicly available, anonymized data; therefore, individual patient consents were not required. All methods were carried out in accordance with relevant guidelines to protect the privacy of patients.

3.1 Evaluation Criteria

Experiments on predicting the lactate trend are evaluated with ten-fold cross-validation. In this cross-validation approach, the dataset is split into ten parts that have an equal number of samples. One part is selected for testing and the rest are used for the training phase. The cross-validation process stops if all parts are used for the testing phase.

Area Under Curve (AUC) score metric is used to assess the classification performance of machine learning algorithms. This score is calculated by drawing a True Positive Rate (Sensitivity) and False Positive Rate (1-Specificity) curve. Then after drawing this curve area under the curve is calculated to assess the classification model. AUC score range is between 0 and 1. AUC score of 1 means that the classification model can distinguish all samples. So, values that are close to 1 indicate better prediction performance.

3.2 Empirical Results

We evaluated RF, NB, J48, LR, and LMT classifiers on the lactate trend prediction task. All experiments were done on a 10-fold CV setting. We conducted our experiments based on three scenarios; sepsis patient`s lactate value has an increasing trend; sepsis patient has a steady lactate value trend and sepsis patient`s lactate value has a decreasing trend.

Table 2. shows classification results for the lactate trend increase scenario. As can be seen from Table 2. LMT and LR algorithms outperformed other classifiers and achieved 0.647 values in terms of AUC. NB comes second. J48 decision tree performed worse when predicting lactate trend increase. LMT algorithm with 4 features (heart rate, oxygen saturation, lactate value before sepsis diagnosis, and time interval) achieved 0.630 in terms of AUC.

Table 3. shows classification results for the constant lactate trend scenario. As can be seen from Table 3. LMT algorithm outperformed other classifiers and achieved 0.803 value in terms of AUC. RF comes second and LR comes third. J48 decision tree performed worse when predicting constant lactate trend. LMT algorithm with 4 features achieved 0.80 in terms of AUC.

Table 4. shows classification results for the lactate trend decrease scenario. As can be seen from Table 4. LMT algorithm outperformed other classifiers and achieved 0.847 value in terms of AUC. RF comes second and LR comes third. J48 decision tree performed worse when predicting constant lactate trend. LMT algorithm with 4 features achieved 0.844 in terms of AUC. According to experimental results, we can say that machine learning models that employ logistic regression architectures overall achieved good results in lactate trend prediction tasks. Also, the LMT algorithm with just four variables achieved a noteworthy prediction performance compared with the LMT algorithm that uses all of the variables. Especially in constant lactate trend and lactate trend decrease scenarios, LMT with 4 features achieved very similar results to the LMT algorithm. We can say that the LMT algorithm with heart rate, oxygen saturation, lactate value before sepsis diagnosis and time interval variables can be effectively used to assess the state of the patient whether it is stable or improving.

The LMT models which are one of the machine learning approaches were the most accurate in predicting serum lactate trends from non-invasive clinical variables of patients with sepsis. In this method the AUC of increasing, constant, and decreasing lactate values were 0.647 (95% CI 0.633-0.661), 0.803 (95% CI 0.798-0.808)), and 0.847 (95% CI 0.841-0.853), respectively.

We observed different rankings of the importance of the variables for predicting lactate trends. For example, previous serum lactate measurement was a significant predictor of change in serum lactate values, followed by oxygen saturation, the time interval between lactate measurements, heart rate, SBP, P/F ratio, age, MBP, and DBP.

Multiple studies are conducted every year on reducing the fatality rate associated with sepsis. Early identification and prediction of patients at risk of adverse outcomes related to sepsis and septic shock are critical to timely intervention. Today, there are many studies on the early predictions of sepsis that rely on determining the presence of infection and organ dysfunction by screening laboratory tests (such as procalcitonin, C-reactive protein, white blood cells, platelets, and lactic acid) [33-36]. Even though they are accurately predicted by laboratory testing, these take time, and treatment is further delayed [34, 37].

Markers of sepsis-induced tissue hypoperfusion include indices of global (i.e., hypotension, tachycardia, oliguria, delayed capillary refill, clouded sensorium, elevated blood lactate, low mixed venous O₂ saturation) and regional hypoperfusion [38]. Changes in vital signs (i.e., heart rate, blood pressure, respiratory rate, oxygen saturation, and body temperature) are seen several hours before a severe adverse event occurrence and clinical deterioration. The Systemic Inflammatory Response Syndrome (SIRS) and qSOFA score (also known as the rapid Sepsis-Related Organ Failure Assessment) are largely based on identifying changes in vital signs. These criteria remain an important clinical tool for the host's systemic response to inflammation, despite the discovery of several biomarkers [39].

Several studies suggest that trend analysis of intermittent vital signs would increase the accuracy of detecting clinical deterioration in general wards and emergency departments [40]. According to a study by Barfod et al. [41] abnormal vital signs (SpO₂, RR, BP, HR, GCS), especially abnormal RR, SpO₂, and GCS, are strong predictors for intensive care unit admission from the emergency department and in-hospital mortality. Today, some researchers are developing sepsis diagnosis and mortality prediction models by analyzing the change in vital signs using machine learning techniques. A machine learning-based sepsis prediction algorithm (InSight) developed by Mao Q et al. [42] provides high sensitivity and specificity for the detection and prediction of sepsis, severe sepsis, and septic shock using only six common vital data acquired in the emergency department, general ward and ICU.

In clinical conditions, the circulatory disorder may be characterized by abnormal hemodynamic parameters such as hypotension and tachycardia, abnormal tissue organ perfusion findings such as decreased urine output and changes in consciousness, and abnormal metabolic parameters such as increased lactate and metabolic acidosis [43].

Hyperlactatemia is common in patients with sepsis, which is a marker of disease severity and a strong predictor of mortality. Sepsis-associated hyperlactatemia may reflect the degree of activation of the stress response (and epinephrine release) [1]. In daily clinical practice, it is accepted that the change in lactate levels over time primarily reflects a change in its production. As hyperlactatemia is often associated with poor circulation, we usually see a decrease in lactate levels in the improved circulation state, and we hypothesize (but cannot prove) decreased production [44]. However, since clearance is significantly reduced in stable septic patient shock states, continued hyperlactatemia may reflect decreased clearance rather than increased lactate production [5]. There is a significant relationship between lactate level and risk of death, and it could be used to risk-stratify patients' disease severity. High lactate level indicates an increased risk of mortality, and it can help identify patients who need additional investigation and monitoring [45].

Empirical results reveal that machine learning approaches that utilize logistic regression functions achieved higher AUC values as compared to others. Due to its robust structure to overfitting, LMT algorithm achieved high AUC values even with just four features. Experimental results also prove that the LMT algorithm can be combined with easily acquirable and routinely collected parameters to predict the lactate trend of sepsis patients. LMT model has high computational complexity due to its hybrid logistic tree structure. If the model is trained with high dimensional data, it can lead to high CPU and memory consumption [46]. To overcome this issue, our approach only uses parameters that have high importance in the lactate trend prediction task. With this proposed approach, a quick and accurate solution based on easily acquirable wearable parameters can be implemented in the ICU setting to assess the trend of lactate value.

Lactate metabolism is affected by many factors, thus predicting its level using machine learning is not easy. Treatment can be tailored according to predicting the lactate trend rather than predicting one single value. Our study suggests that lactate change can be predicted with a suboptimal performance by using machine learning models that use hemodynamic and respiratory parameters of patients. Further clinical studies are needed to evaluate its usability and utility in clinical practice.

By adding more lactate-related parameters to the dataset, the performance of deep learning methods, which is a branch of machine learning, can be examined. Deep learning structures that have a good performance, such as LSTM, and CNN can be combined with LMT to form a hybrid structure and be used in predicting lactate trends

The need to find a better way to predict patients’ survival is still ongoing. Machine learning is gaining more importance and attention as the clinical outcomes are well correlated with the predictions of the systems used. Non-invasive and less costly approaches with accurate estimations of the patients are preferred by clinicians. Predicting the lactate trend, in other words, the state of sepsis patient whether it is stable or improving, in the ICU by LMT algorithm which uses heart rate, oxygen saturation, lactate value before sepsis diagnosis and time interval variables can be done effectively.

AUC: Area under the curve

BIDMC: Beth Israel Deaconess Medical Center

BP. Blood pressure

CFS: Correlation-based feature selection

CNN: Convolutional neural network

CPU: Central Processing Unit

CV Setting: Cross-Validation Setting

DBP: Diastolic blood pressure

FiO₂: Fraction of inspired oxygen

GCS: Glasgow Coma Scale

HR: Heart rate

ICU: Intensive care unit

LMICs: Low- and middle-income countries (

LMT: Logistic Model Tree

LR: Logistic Regression

LSTM: Long short-term memory

MBP: Mean blood pressure

MMIC: The Medical Information Mart for Intensive Care

NB: Naïve Bayes

P/F ratio: Pao₂/Fio₂: The ratio of arterial oxygen partial pressure (PaO2 in mmHg) to fractional inspired oxygen (FiO₂ expressed as a fraction, not a percentage)

PaO₂:Partial pressure of oxygen in arterial blood

qSOFA: Quick Sepsis-Related Organ Failure Assessment)

RF: Random Forest

RR: Respiratory rate

SBP: Systolic blood pressure

SIRS: Systemic Inflammatory Response Syndrome

SpO₂: Peripheral oxygen saturation

SSC: The Surviving Sepsis Campaign

STROBE: Strengthening the Reporting of Observational Studies in Epidemiology

Author contributions

Mustafa Kemal Arslantas (MKA), Tunc Asuroglu (TA), Reyhan Arslantas (RA), Emin Pashazade (EP), Gulbin Tore Altun (GTA), Pelin Corman Dincer (PCD), and Alper Kararmaz (AK) conceived and planned the study. They were mainly responsible for the design of the study. Acquisition of data: RA, EP. Analysis and interpretation of data: MKA, TA. Drafting of the article: TA, RA, GTA, PCD. Critical manuscript revision: MKA, AK. Review of the submitted version of the manuscript: MKA, TA, RA, EP, GTA, PCD. Study supervision: MKA, AK. All authors gave final approval of the submitted manuscript. All authors have also provided an intellectual contribution to the manuscript.

Funding: None

Conflict of interest: The authors declare that there are no conflicts of interest

Ethical approval: The project is approved by the institutional review boards of MIT (Cambridge, MA, USA) and BIDMC (Boston, MA, USA) and was granted a waiver of informed consent.

Garcia-Alvarez M, Marik P, Bellomo R. Sepsis-associated hyperlactatemia. Crit Care. 2014;18(5):503. doi: 10.1186/s13054-014-0503-3.
Ryoo SM, Lee J, Lee YS, Lee JH, Lim KS, Huh JW, et al. Lactate Level Versus Lactate Clearance for Predicting Mortality in Patients With Septic Shock Defined by Sepsis-3. Crit Care Med. 2018;46(6):e489-e95. doi: 10.1097/CCM.0000000000003030.
Masyuk M, Wernly B, Lichtenauer M, Franz M, Kabisch B, Muessig JM, et al. Prognostic relevance of serum lactate kinetics in critically ill patients. Intensive Care Med. 2019;45(1):55-61. doi: 10.1007/s00134-018-5475-3.
Liu Z, Meng Z, Li Y, Zhao J, Wu S, Gou S, et al. Prognostic accuracy of the serum lactate level, the SOFA score and the qSOFA score for mortality among adults with Sepsis. Scand J Trauma Resusc Emerg Med. 2019;27(1):51. doi: 10.1186/s13049-019-0609-3.
Nguyen HB, Rivers EP, Knoblich BP, Jacobsen G, Muzzin A, Ressler JA, et al. Early lactate clearance is associated with improved outcome in severe sepsis and septic shock. Crit Care Med. 2004;32(8):1637-42. doi: 10.1097/01.ccm.0000132904.35713.a7.
Claridge JA, Crabtree TD, Pelletier SJ, Butler K, Sawyer RG, Young JS. Persistent Occult Hypoperfusion Is Associated with a Significant Increase in Infection Rate and Mortality in Major Trauma Patients. Journal of Trauma and Acute Care Surgery. 2000;48(1):8.
Evans L, Rhodes A, Alhazzani W, Antonelli M, Coopersmith CM, French C, et al. Surviving Sepsis Campaign: International Guidelines for Management of Sepsis and Septic Shock 2021. Crit Care Med. 2021;49(11):e1063-e143. doi: 10.1097/CCM.0000000000005337.
Pan J, Peng M, Liao C, Hu X, Wang A, Li X. Relative efficacy and safety of early lactate clearance-guided therapy resuscitation in patients with sepsis: A meta-analysis. Medicine (Baltimore). 2019;98(8):e14453. doi: 10.1097/MD.0000000000014453.
Wolf A, Renehan K, Ho KKY, Carr BD, Chen CV, Cornell MS, et al. Evaluation of Continuous Lactate Monitoring Systems within a Heparinized In Vivo Porcine Model Intravenously and Subcutaneously. Biosensors (Basel). 2018;8(4). doi: 10.3390/bios8040122.
Timsit JF, Ruppe E, Barbier F, Tabah A, Bassetti M. Bloodstream infections in critically ill patients: an expert statement. Intensive Care Med. 2020;46(2):266-84. doi: 10.1007/s00134-020-05950-6.
Jakacka N, Snarski E, Mekuria S. Prevention of Iatrogenic Anemia in Critical and Neonatal Care. Adv Clin Exp Med. 2016;25(1):191-7. doi: 10.17219/acem/32065.
Kwizera A, Adhikari NKJ, Angus DC, Dondorp AM, Dunser MW, Festic E, et al. Recognition of Sepsis in Resource-Limited Settings. In: Dondorp AM, Dunser MW, Schultz MJ, editors. Sepsis Management in Resource-limited Settings. Cham (CH)2019. p. 69-84.
Mamandipoor B, Yeung W, Agha-Mir-Salim L, Stone DJ, Osmani V, Celi LA. Prediction of blood lactate values in critically ill patients: a retrospective multi-center cohort study. J Clin Monit Comput. 2021. doi: 10.1007/s10877-021-00739-4.
Johnson A, Bulgarelli L, Pollard T, Horng S, Celi LA, Mark R: MIMIC-IV (version 1.0). PhysioNet. https://doi.org/10.13026/s6n6-xd98 (2021). Accessed.
von Elm E, Altman DG, Egger M, Pocock SJ, Gotzsche PC, Vandenbroucke JP, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. J Clin Epidemiol. 2008;61(4):344-9. doi: 10.1016/j.jclinepi.2007.11.008.
Stevens LM, Mortazavi BJ, Deo RC, Curtis L, Kao DP. Recommendations for Reporting Machine Learning Analyses in Clinical Research. Circ Cardiovasc Qual Outcomes. 2020;13(10):e006556. doi: 10.1161/CIRCOUTCOMES.120.006556.
Singer M, Deutschman CS, Seymour CW, Shankar-Hari M, Annane D, Bauer M, et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA. 2016;315(8):801-10. doi: 10.1001/jama.2016.0287.
Chaudhuri AK, Sinha D, Banerjee DK, Das A. A novel enhanced decision tree model for detecting chronic kidney disease. Network Modeling Analysis in Health Informatics and Bioinformatics. 2021;10(1). doi: 10.1007/s13721-021-00302-w.
Sharma A, Mishra PK. Performance analysis of machine learning based optimized feature selection approaches for breast cancer diagnosis. International Journal of Information Technology. 2021;14(4):1949-60. doi: 10.1007/s41870-021-00671-5.
Bužić D, Dobša J. Lyrics classification using Naive Bayes. 2018 41st International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO)2018. p. 1011-5.
Bhat P, Malaganve P. Effect of J48 and LMT Algorithms to Classify Movies in the Web—A Comparative Approach. Singapore: Springer Singapore; 2021. p. 547-53.
Aşuroğlu T, Açıcı K, Berke Erdaş Ç, Kılınç Toprak M, Erdem H, Oğul H. Parkinson's disease monitoring from gait analysis via foot-worn sensors. Biocybernetics and Biomedical Engineering. 2018;38(3):760-72. doi: https://doi.org/10.1016/j.bbe.2018.06.002.
Ilyas H, Ali S, Ponum M, Hasan O, Mahmood MT, Iftikhar M, et al. Chronic kidney disease diagnosis using decision tree algorithms. BMC Nephrol. 2021;22(1):273. doi: 10.1186/s12882-021-02474-z.
Rahmayanti N, Pradani H, Pahlawan M, Vinarti R. Comparison of machine learning algorithms to classify fetal health using cardiotocogram data. Procedia Computer Science. 2022;197:162-71. doi: https://doi.org/10.1016/j.procs.2021.12.130.
Simsekler MCE, Alhashmi NH, Azar E, King N, Luqman R, Al Mulla A. Exploring drivers of patient satisfaction using a random forest algorithm. BMC Med Inform Decis Mak. 2021;21(1):157. doi: 10.1186/s12911-021-01519-5.
Açıcı K, Erdaş ÇB, Aşuroğlu T, Toprak MK, Erdem H, Oğul H. A Random Forest Method to Detect Parkinson’s Disease via Gait Analysis. Cham: Springer International Publishing; 2017. p. 609-19.
Aşuroğlu T, Oğul H. A deep learning approach for sepsis monitoring via severity score estimation. Computer Methods and Programs in Biomedicine. 2021;198:105816. doi: https://doi.org/10.1016/j.cmpb.2020.105816.
Choudhury A, Gupta D. A Survey on Medical Diagnosis of Diabetes Using Machine Learning Techniques. Singapore: Springer Singapore; 2019. p. 67-78.
Kim A, Song Y, Kim M, Lee K, Cheon JH. Logistic regression model training based on the approximate homomorphic encryption. BMC Medical Genomics. 2018;11(4):83. doi: 10.1186/s12920-018-0401-7.
Nazari E, Aghemiri M, Avan A, Mehrabian A, Tabesh H. Machine learning approaches for classification of colorectal cancer with and without feature selection method on microarray data. Gene Reports. 2021;25:101419. doi: https://doi.org/10.1016/j.genrep.2021.101419.
Jha SK, Pan Z, Elahi E, Patel N. A comprehensive search for expert classification methods in disease diagnosis and prediction. Expert Systems. 2019;36(1). doi: 10.1111/exsy.12343.
Tien Bui D, Shahabi H, Omidvar E, Shirzadi A, Geertsema M, Clague JJ, et al. Shallow Landslide Prediction Using a Novel Hybrid Functional Machine Learning Algorithm. Remote Sensing. 2019;11(8):931.
Opal SM, Wittebole X. Biomarkers of Infection and Sepsis. Crit Care Clin. 2020;36(1):11-22. doi: 10.1016/j.ccc.2019.08.002.
Herzum I, Renz H. Inflammatory markers in SIRS, sepsis and septic shock. Curr Med Chem. 2008;15(6):581-7. doi: 10.2174/092986708783769704.
Ljungström L, Pernestig A-K, Jacobsson G, Andersson R, Usener B, Tilevik D. Diagnostic accuracy of procalcitonin, neutrophil-lymphocyte count ratio, C-reactive protein, and lactate in patients with suspected bacterial sepsis. PLOS ONE. 2017;12(7):e0181704. doi: 10.1371/journal.pone.0181704.
Arslantas MK, Arslantas R, Dincer PC, Altun GT, Kararmaz A. Prognostic Value of the Lactate-Albumin Difference in Predicting 30-Day Mortality in Critically Ill Patients. Shock. 2021;56(1):150-1. doi: 10.1097/SHK.0000000000001613.
Seymour CW, Gesten F, Prescott HC, Friedrich ME, Iwashyna TJ, Phillips GS, et al. Time to Treatment and Mortality during Mandated Emergency Care for Sepsis. N Engl J Med. 2017;376(23):2235-44. doi: 10.1056/NEJMoa1703058.
Zanotti Cavazzoni SL, Dellinger RP. Hemodynamic optimization of sepsis-induced tissue hypoperfusion. Crit Care. 2006;10 Suppl 3:S2. doi: 10.1186/cc4829.
Gul F, Arslantas MK, Cinel I, Kumar A. Changing Definitions of Sepsis. Turk J Anaesthesiol Reanim. 2017;45(3):129-38. doi: 10.5152/TJAR.2017.93753.
Brekke IJ, Puntervoll LH, Pedersen PB, Kellett J, Brabrand M. The value of vital sign trends in predicting and monitoring clinical deterioration: A systematic review. PLoS One. 2019;14(1):e0210875. doi: 10.1371/journal.pone.0210875.
Barfod C, Lauritzen MM, Danker JK, Soletormos G, Forberg JL, Berlac PA, et al. Abnormal vital signs are strong predictors for intensive care unit admission and in-hospital mortality in adults triaged in the emergency department - a prospective cohort study. Scand J Trauma Resusc Emerg Med. 2012;20:28. doi: 10.1186/1757-7241-20-28.
Mao Q, Jay M, Hoffman JL, Calvert J, Barton C, Shimabukuro D, et al. Multicentre validation of a sepsis prediction algorithm using only vital sign data in the emergency department, general ward and ICU. BMJ Open. 2018;8(1):e017833. doi: 10.1136/bmjopen-2017-017833.
Bakker J, Postelnicu R, Mukherjee V. Lactate: Where Are We Now? Crit Care Clin. 2020;36(1):115-24. doi: 10.1016/j.ccc.2019.08.009.
Vincent JL, Bakker J. Blood lactate levels in sepsis: in 8 questions. Curr Opin Crit Care. 2021;27(3):298-302. doi: 10.1097/MCC.0000000000000824.
Villar J, Short JH, Lighthall G. Lactate Predicts Both Short- and Long-Term Mortality in Patients With and Without Sepsis. Infect Dis (Auckl). 2019;12:1178633719862776. doi: 10.1177/1178633719862776.
Lee S, Jun C-H. Fast incremental learning of logistic model tree using least angle regression. Expert Systems with Applications. 2018;97:137-45. doi: https://doi.org/10.1016/j.eswa.2017.12.014.

Table 1 Patient Characteristics (N = 18653)
Variable	Lactate Trend			P-value²
Variable	Decrease N = 3313 (18%)¹	Increase N = 1328 (7.1%)¹	Constant N = 14012 (75%)¹	P-value²
Age (y)	65 (15)	65 (15)	65 (15)	0.015
Heart Rate (bpm)	89 (18)	95 (20)	89 (19)	<0.001
Systolic Blood Pressure (mmHg)	111 (14)	106 (15)	110 (15)	<0.001
Mean Blood Pressure (mmHg)	89 (18)	95 (20)	89 (19)	<0.001
Diastolic Blood Pressure (mmHg)	76 (10)	73 (11)	75 (10)	<0.001
Blood Oxygen Saturation Level (%, SpO₂)	98.35 (1.94)	97.75 (2.26)	97.89 (2.17)	<0.001
PaO₂/FiO₂ Ratio (P/F Ratio)	240 (118)	214 (119)	228 (114)	<0.001
Initial Lactate Value (mmol/L)	4.68 (1.66)	3.74 (1.97)	2.67 (1.57)	<0.001
Time Interval (min)	197 (81)	188 (82)	186 (84)	<0.001
¹ Mean (SD)
² Kruskal-Wallis rank sum test

Table 2

Classification results for lactate trend increase

Classification Model	AUC (95 % CI)
RF	0.628 (0.614-0.642)
NB	0.637 (0.623-0.651)
J48 decision tree	0.555 (0.54-0.57)
LR	0.647 (0.633-0.661)
LMT	0.647 (0.633-0.661)
LMT (with 4 features)	0.630 (0.616-0.644)

Table 3

Classification results for constant lactate trend

Classification Model	AUC (95 % CI)
RF	0.794 (0.789-0.799)
NB	0.779 (0.774-0.784)
J48 Decision Tree	0.726 (0.721-0.731)
LR	0.792 (0.787-0.797)
LMT	0.803 (0.798-0.808)
LMT (with 4 features)	0.80 (0.795-0.805)

Table 4

Classification results for lactate trend decrease

Classification Model	AUC (95 % CI)
RF	0.842 (0.836-0.848)
NB	0.822 (0.816-0.828)
J48 Decision Tree	0.751 (0.743-0.759)
LR	0.826 (0.82-0.832)
LMT	0.847 (0.841-0.853)
LMT (with 4 features)	0.844 (0.838-0.85)

No competing interests reported.

Download PDF

Version 2

posted

You are reading this latest preprint version

Using machine learning methods to predict the lactate trend of sepsis patients in the ICU

Status:

Version 2

Abstract

Figures

1. Introduction

2. Methods

2.1. Data sources

2.2 Study design

2.3 Deﬁnition of outcomes

2.5 Variable selection

2.5 Proposed Machine Learning Framework

2.6 Selected Classifiers for Proposed Framework

Ethics approval and consent to participate

3. Results

3.1 Evaluation Criteria

3.2 Empirical Results

4. Discussion

5. Conclusion

Abbreviations

Declarations

References

Tables

Additional Declarations

Status:

Version 2