1.1 Introduction/History
When one thinks of machine learning and artificial intelligence in healthcare, the overwhelming thought is about their applicability in radiology for diagnosing pathologies through chest X-rays/computed tomography or their applicability in dermatology for diagnosing pathologies using skin images. These are logical applications, because we expect that by seeing numerous normal and pathological images, the machine/neural network can decipher the salient features that define each pathology and then use these features to predict pathology in unseen images. The above applications are also the first to come to mind because significant progress has been made in these areas. For example, in May 2019, Google developed a deep learning algorithm for the detection of lung cancer by computed tomography. The deep learning model achieved an area under the receiver operating curve of 94.4% and the model outperformed six radiologists in the diagnosis while at the same time having an absolute reduction of 11% in false positives and 5% in false negatives [1].
Often anesthesiology is not associated with machine learning because the use of images to detect pathology plays a much smaller role in this field than in other healthcare fields. However, machine learning encompasses far more than just image analysis. Machine learning involves a system that is able to perform a task without being explicitly programmed but rather based on algorithms, statistical models, regression, and/or pattern recognition. In addition, the machine learning system is able to learn and improve from experience. This experience is often gained through data, which could include tabular data or pixilated data. In anesthesia, there is a plethora of tabular data. This data could be individual case data, which may consist of variables including gender, age, height, weight, medications, allergies, heart rate, respiratory rate, blood pressure, and temperature. These types of variables, especially those that are monitored and can fluctuate, such as vital signs, can be provided to a machine learning model in combination with another variable of interest such as hypotension. With enough information, the machine learning model can analyze the data in order to develop an association between the dependent variables and the independent outcome variable. After the creation of this type of regression model, the neural network can be provided with new cases/data and can predict the outcome variable based on the association it developed from the training data. In addition, as the neural network is provided with more data, it will continue to update the association between the variables it developed before. As such, this model is performing the task of predicting an independent variable without any explicit algorithm or brute force mechanism being implemented by a user, and the model is learning and improving from experience. The above theoretical example provides some of the underlying information as to how machine learning can be utilized in anesthesiology. Later in this chapter, concrete examples of machine learning in action in anesthesia are provided.
Automation in the field of anesthesiology is not new. One of the earliest recorded implementations of automation includes Bickford’s efforts in 1950 [2]. Bickford recognized a relationship between the summated value of brain potentials and the degree to which the central nervous system was depressed by the anesthetic agent. With this information, Bickford designed an apparatus to automate the maintenance of anesthesia. Since 1950, new variables and new combinations of these variables have been identified that more accurately correlate to the depth of anesthesia. Integrating this information into machines led the Food and Drug Administration to approve Sedasys computer-assisted personalized sedation [3]. This machine was designed to integrate propofol delivery to the patient with monitoring of the patient through measures including pulse oximetry, electrocardiogram, and patient response to auditory and tactile stimuli [4]. There were limitations to its use: only for mild to moderate sedation and only for either a colonoscopy or an esophagogastroduodenoscopy [5]; however, the Sedasys machine is a great example of how automation and the use of algorithm have been used to aid anesthesiologists.
In general, much of the automation that has been developed in anesthesiology requires a programmer to input an algorithm that the machine uses in order to solve a task. The identification of this algorithm is very much still dependent on a human. Furthermore, with new information, it is up to the human to update the algorithm, which often takes time and leads to delay in implementation. This can be compared to machine learning in which the neural network identifies the algorithm and continuously updates it as new data are provided.
1.2 Machine Learning Implementation in Anesthesia
In the last few years, several applications of machine learning have been implemented in anesthesia including preoperative, perioperative, and postoperative applications.
1.2.1 Hypotension
Two research studies have involved the prediction of hypotension after anesthesia induction. Hatib and colleagues used the dependent variable, high-fidelity arterial pressure waveform, in order to predict the onset of hypotension prior to its actual onset [6]. For training, 545,959 minutes of arterial waveform recording and 25,461 episodes of hypotension were included. In addition, for validation, 33,326 minutes of arterial waveform recording and 1,923 episodes of hypotension were included. Even though on the surface it seems that there was only one dependent variable being analyzed, in fact, over 3,000 individual features were selected from the arterial pressure waveform and each one was run in the analysis in order to generate a receiver operating curve. Afterward, 51 features were selected based on having an area under the curve (AUC) on the receiver operative curve of greater than 0.85. These 51 features/variables with their reciprocal and squared terms and in combination generated a total of over 2.6 million total variables. The model was then able to calculate linear and nonlinear relationships between these variables and hypotension. In the end, the model used 3,022 individual features and achieved a sensitivity, specificity, and AUC of 88%, 87%, and 0.95, respectively, when predicting a hypotensive event 15 minutes before its onset. In predicting a hypotensive event 5 minutes before its onset, these values increased to 92%, 92%, and 0.97, respectively. In addition, Kendale and colleagues implemented machine learning in the prediction of postinduction hypotension [7]. However, in this analysis, the dependent variables included additional characteristics, such as demographics, preoperative medications, medical comorbidities, induction medications, and intraoperative vital signs. In addition, eight different supervised machine-learning classification techniques (logistic regression, support vector machines, naïve Bayes, K-nearest neighbor, linear discriminant analysis, random forest, neural network, and gradient boosting machine) were compared for their performance. Gradient boosting was found to have the highest AUC and the final model had an AUC of 0.77. The better performance of gradient boosting aligns with the type of data as it performs well with class imbalance. Finally, the dependent variables were compared for their importance in the prediction of postinduction hypotension. The top five in importance were first mean arterial pressure, age, BMI, mean peak inspiratory pressure, and maximum sevoflurane concentration. In addition, the model uncovered a high importance of the medications- levothyroxine and bisphosphonates, which would likely not have been expected. Predicting hypotension is a very challenging task and if machine learning and artificial intelligence continue to progress as in the above two studies, anesthesiologists will be less likely to be caught off guard and therefore will be able to provide treatment more quickly and more effectively, possibly even providing preventive treatment in some cases.
1.2.2 Hypoxia
Another study performed a gradient boosting machine model in order to predict hypoxemia during surgery [8]. Some of the features/variables included were patient demographics (e.g. BMI), BP, respiration rate, O2 flow, medications, SpO2, and time since the start of the procedure. The machine learning system was utilized for initial risk prediction and for real-time hypoxemia prediction. In both cases, the combination of an anesthesiologist and the machine learning system led to a statistically significant (P < 0.0001) increase in the AUC. With the amount of data that is available prior to surgery and during surgery, machine learning can provide clinicians with a support system to help them predict life-threatening cases before they occur.
1.2.3 Depth of Anesthesia
Monitoring the depth of anesthesia during a surgical procedure is crucial for patient safety and patient comfort. A high depth of anesthesia can lead to delayed recovery time, hypotension, and decreased perfusion to vital organs. On the other hand, a low depth of anesthesia can lead to post-traumatic stress. Saadeh and colleagues designed a machine learning-based EEG processor that estimated the depth of anesthesia [9]. Six features were extracted from the EEG (spectral edge frequency, beta ratio, and four bands of spectral energy) and were then used in the machine learning system to classify the depth of anesthesia as deep, moderate, light, or awake. During testing, the system achieved 92.2% accuracy in classifying the depth of anesthesia, again highlighting the benefit of machine learning in the operating room.
1.2.4 Mortality Prediction
Predicting mortality is often a task left up to a higher power. In...