Decision support system to improve postoperative discharge: A novel multi-class classification approach
Introduction
Currently, physicians are let alone to assess the clinical status of the patients postoperatively based on several factors to decide on where they should be sent to next after undergoing a surgical intervention. Such a postoperative decision making may affect the incidence of postoperative complications, impairing the survival in some patients and the management of the hospital resources (financial resources, e.g., equipment, hospital beds, etc.; human resources, e.g., clinicians and nurses, etc.). Resources in the intensive care unit (ICU) are limited and, therefore, it is vital to prioritise them for patients who need critical medical attention. In a general ward, hospital resources are very limited too, e.g., hospital beds, thus an appropriate management of such resources is required. The high volume of patients under observation and the induced stress by working for long shifts often result in work-related fatigue in clinicians, which may impair the objectivity in their decision making, further complicating patient management and, in some cases, increasing mortality [1].
Hypothermia is a condition that occurs whenever the patient's core body temperature is lower than the physiological temperature (36°C). Postoperative hypothermia is common. as, during surgery, administration of anaesthetics and environmental factors can lower a patient's body temperature [26]. Hypothermia may lead to increased oxygen consumption, increased patient discomfort and physiological instability [27]. Persistent hypothermia is associated with increased mortality in both patients undergoing cardiac [22] and non-cardiac surgery [9]. Thus, early detection and prevention of hypothermia may improve postoperative recovery. Maintaining a physiological body temperature is not only essential for survival [22] but can also help reduce the length of hospital stay, the incidence of surgical site infections, postoperative blood transfusions, pressure ulcers, as well as patient discomfort and, ultimately, mortality [20]. Nevertheless, hypothermia is often undiagnosed and left untreated [3].
Postoperative discharge decision making, as any clinical decision-making processes, is complex, since it involves several variables that may be associated with each other and with different levels of significance, such as demographics, physical status, body temperature, blood pressure, level of discomfort upon discharge [24], etc. Most of the current systems available in hospitals to aid postoperative discharge decision-making are based on statistical tools, which, however, lack the capability of capturing the underlying patterns of such highly nonlinear and dimensional data to make accurate and reliable predictions.
Machine Learning (ML) has been successfully used to improve prediction of postoperative discharge, thus showing potential in improving management of hospital resources and reducing the incidence of postoperative complications [1], e.g., pulmonary-related ones, such as bronchitis, pneumonia, pleural effusions and prolonged mechanical ventilation leading to acute respiratory failure [2]. As compared to statistical-based methods, such as principal component analysis (PCA), ML has been proven to capture the underlying patterns of clinical data further [13].
Postoperative discharge decision making inherently involves a multi-class classification problem, wherein there may be potentially multiple clinical outcomes associated with the patient status, baseline and postoperative characteristics, thus resulting in more than two potential output classes, e.g., further to undergoing surgery, patients may be discharged home or, alternatively, be sent to a general ward if further observation and assistance are required during their recovery, or, should the conditions of the patients deteriorate critically following surgery, sent to the ICU. ML applies the learnt relationship from the input data used for training, once validated on a separate dataset (the cross-validation data), to predict it on additional data samples, namely the testing set.
As Hansen and McDonald [7] had pointed out, there is still no objective method for optimising hyperparameters of Machine Learning-based classifiers. Currently, parameterisation of such learning parameters heavily relies on the subjective experience of Machine Learning engineers and heuristic methods, which are strictly task- and data-specific and, therefore, not generalisable. Conversely, Genetic Algorithm (GA)-based unconstrained optimisation enables to select optimal hyperparameters [19] that yield consistently higher classification performance than heuristic and statistical methods [7], thus showing potential in improving the accuracy and reliability of ML-based multi-class classification methods for aiding prediction of postoperative discharge.
In an attempt of dealing with such uncertainties on postoperative data, linguistic data were represented as fuzzy numbers and principal component analysis (PCA) was used for feature reduction in an attempt of improving classification performance in previous studies [11].
Luuka [11] implemented a fuzzy- and (PCA)-based similarity classifier and managed to predict the most appropriate recovery area (home, general hospital ward or ICU) in the same dataset of 90 patients developed by Summers et al. [24] with 66.22% testing classification accuracy (ACC). Forghani and Yazdi [5] proposed a more complex algorithm, the fuzzy min–max neural network with symmetric margin (FMNWSM) and applied it to classify the same data set [24], achieving 72.99% of ACC, as compared to 74.75% with Support Vector Machine (SVM), but with a considerably lower computational cost than SVM (2.1 ms for the FMNWSM algorithm against 4583 ms for the SVM). Although FMNWSM converges faster than conventional fuzzy min-max neural networks due to the lack of special nodes in overlapped regions [5], its architecture is too complex for it to be used in a clinical setting, considering that its accuracy (72.99%) is not clinically valuable.
ML has been successfully applied to avoid hypothermia by predicting the most appropriate recovery area for each patient (home, general ward or ICU) [8]. Hsieh et al. [8] used a particle sward optimisation (PSO)-based fuzzy classifier to retrieve the crisp rules from the same postoperative data [24], where such rules were used to aid prediction of clinical decision making, which was 84% accurate.
Using 10-fold cross validation (CV), Abuaqel et al. [1] applied SVM and Artificial Neural Networks (ANN) in Weka to predict postoperative discharge on the same data set of interest [24], obtaining 88.54% and 82.81% of classification accuracy respectively.
To the best of authors’ knowledge, the M-SVM (SVM for multi-class classification) used by Abuaqel et al. [1] is the best performing classifier on this data set [24] to date.
Temperature control post-surgery can aid early recovery, thus increasing patient comfort, decreasing shivering and associated oxygen saturation, facilitating physiological stability and leading to better postoperative outcomes [27]. We hypothesise that ML can assist physicians and nurses in making an evidence-based informed decision on where patients should be sent to next after surgery, thus optimising clinical outcome and management of hospital resources, also improving survival in some cases. Improved physiological stability can reduce the length of hospitalisation in an expensive recovery room substantially, thus also saving additional costs to the hospital, insurance companies and, in some cases, to the patient themselves [27].
Therefore, such expert systems using ML can help prevent hypothermia, thus also solving the problem of information overload, typical in healthcare as described in Sections1.1 and 1.2, and, therefore, improving the quality of patient care and clinical outcomes [27]. Nevertheless, the underlying engineering process for devising a ML-expert system with predictive capability on selecting the most appropriate recovery area post-surgery for each patient is still either too complex to be applied in a clinical setting [5], [11] or inadequately described [8].
Considering that the lack of clinical translation of previous research findings [5], [8], [11], has left the clinical challenges discussed in 1.1. still open, this study aims to develop and validate a novel hybrid algorithm that can accurate and reliably predict postoperative discharge in case of unbalanced classes, typical in healthcare data of this type.
Section snippets
Data pre-processing and encoding methods
Postoperative discharge-related data from the University California-Irvine (UCI) Machine Learning repository on ninety (N = 90) patients were used [24]. Since hypothermia is a concern post-surgery, the input features are related to body temperature measurements.
The eight input features for classification are the following:
- 1.
Patient's internal temperature (L-CORE in°C), high (>37), mid (>=36 and <=37), low (<36);
- 2.
Patient's surface temperature (L-SURF in °C), high (>36.5), mid (>=36.5 and <=35), low
GA-based optimisation of machine learning hyperparameters to improve classification performance
GA-based optimisation was applied to select the optimal hyperparameters ν and α in the M-LSVM used for classification in this study. Considering the number of input features (N = 8), 8-fold CV was selected [18] and 8 hidden neurons were used in the hidden layer of the M-MLP. Initially, the M-LSVM was applied for classifying all input data features (N = 8) with default hyperparameters, i.e., ν = 1/size(data,1), α = 1.9/ν [12]. Further to performing a GA-based optimisation to optimise ν and α in
The advantage of ‘controlled’ All-vs-All for multi-class classification of data on postoperative discharge
The M-MLP, M-SVM and M-LSVM were first used individually for predicting the optimal recovery area based on all input postoperative discharge-related data (N = 8); the M-MLP could reach 92.31% of testing classification accuracy (AUC = 0.65, 95% CI: 0.61–0.68) (Table 2), as also demonstrated in a previous study [18], which satisfies the hypothesis whereby the MLP is a suitable algorithm for classifying highly dimensional and nonlinear data. However, the decrease in specificity (60%, 95%CI
Conclusion
Following a comprehensive analysis of the classification performance of the proposed hybrid model (GA-M-LSVM) against the M-MLP, the M-SVM and the M-LSVM, as well as the best performing ML-based classifiers from published studies, the GA-M-LSVM proved to be the most accurate, reliable and computationally the fastest ML-based classifier for predicting the optimal recovery area on a patient-specific basis, using postoperative discharge-related data. Therefore, the GA-M-LSVM can aid prediction of
Acknowledgment
The authors L.P. and N.R. would like to thank the University of Auckland for giving them the opportunity to carry out their Ph.D. research projects. The authors L.P. and N.R. would like to thank the University of Auckland Rehabilitative Technologies Association (UARTA) and MedIntellego®, for giving them the chance of developing this collaborative research work.
Declaration of interest
The authors declare no conflicts of interest.
Contributors
All authors directly participated in the planning, execution and analysis in the study. All authors also approved the final version of the manuscript, and this submission for possible publication in Knowledge-Based Systems.
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
References (27)
- et al.
Postoperative complications: does intensive care unit staff nursing make a difference
Heart Lung
(2002) - et al.
Evaluation of simple performance measures for tuning SVM hyperparameters
Neurocomputing
(2003) - et al.
Receiver operating characteristic curve analysis of clinical risk models
Ann. Thorac. Surg.
(2001) - et al.
Prediction models aided postoperative decision making based on neural network and support vector machines
- et al.
Accidental hypothermia: a community hospital perspective
Postgrad. Med.
(1981) - et al.
Fuzzy Min–Max Neural Network for Learning a Classifier with Symmetric Margin
Neural Process. Lett.
(2015) - et al.
Some experimental evidence on the performance of GA-designed neural networks
J. Exp. Theoretical Artif. Intell.
(2001) - et al.
Prediction of postoperative recovery based on a computational rules extractor
- et al.
Postoperative hypothermia and patient outcomes after major elective non-cardiac surgery
Anaesthesia
(2013) - et al.
A sequential dual method for large scale multi-class linear SVMs