Harnessing Artificial Intelligence in the ICU
The medical decision making process is one that relies heavily on the incorporation of various data to accurately diagnose and to effectively treat a patient. As technology continues to evolve, so has the amount of data available for the clinician to use in the medical decision making process. In order to consolidate important variables that predict disease and their treatments, algorithms and decision making trees have been developed to help guide the clinician in the diagnostic and therapeutic processes. Despite their successes, these strategies, along with the human limitations of data analysis, struggle to incorporate the vast amounts of static and dynamic data from a patients’ electronic health records (EMR) and monitoring devices, leaving lots of valuable information out of the decision making process.
Recent technologies known as machine learning (ML) have continued to develop over the past decade in order to better augment medical decision making. Machine learning is able to teach computer software to identify and predict outcomes with high accuracy and precision through pattern recognition, variable association, and numerous examples that either fit or reject algorithm parameters. In the context of postoperative cardiac surgery, ML is able to incorporate several real-time and dynamic data points such as telemetry wave form data derived from sources like electrocardiograms (ECG), intensive care unit (ICU) mechanical ventilator settings, left ventricular assist device (LVAD) settings, pulse oximetry, blood pressure, and other clinically significant variables associated with certain disease states. This data can be collected via standard of care devices and then transmitted to a secure data collection center where it is organized in a machine learning ecosystem to be analyzed for predictive variables in order to assist the physician in making real-time care decisions to help improve patient outcomes.
Prediction of Ventilation Weaning following Cardiac Surgery
Successful ventilation weaning is an important part of the extubation and recovery process following cardiac surgery, especially given the frequent use of heart and lung machines called cardiac bypass. Patients who are successfully weaned from ventilation early are known to have less complications, lower mortality, and lower cost of care compared to patients who require prolonged ventilation.3; 4 Aggressive weaning attempts can be performed to maximize the amount of patients removed from ventilation support, but premature extubation can lead to the inability to reestablish ventilatory support, airway trauma, and failure to oxygenate leading to ischemia. 4
Cardiac complications requiring the reestablishment of mechanical ventilation include ischemia, pulmonary edema in patients with cardiac dysfunction, and volume overload in patients with heart failure. Overall, cardiac dysfunction is estimated to be contributory to ventilation weaning difficulty in 20-60% of patients.
Current ventilation weaning protocol in mechanically ventilated adults typically includes performing a spontaneous breathing test (SBT). If a patient is unable to pass a SBT, they are considered to be difficult to wean. In these circumstances, evaluating and treating the cause of weaning difficulty is critical. However, a recent study has shown that 25-40% of patients who pass a SBT are still unable to get off the ventilator. This highlights the need for a more robust and accurate method of predicting the likelihood of successful ventilation weaning.
The simultaneous accumulation and analysis of all ventilation, chart, and dynamic patient data in predicting successful weaning is simply not feasible for a human-being. With the emergence of machine learning (ML), real-time analysis of all patient data, including data that is actively being collected by the mechanical ventilator or other devices, may be incorporated into algorithms to predict the success rate of extubation and even indicate which SBT method to use (e.g. T-Piece versus pressure support). With the information provided from ML, the treating physician will have the opportunity to optimize individual ventilation strategies and extubate as soon as is optimal.
With ML, it is feasible to teach predictive models to identify variables associated with not only a successful SBT, but actual successful ventilation weaning in the cardiac postoperative care setting. A study done by Totonchi et al. found that a history of hypertension, chronic obstructive pulmonary disease (COPD), chronic kidney disease, and previous endocarditis were all predictive of delayed extubation in patients undergoing cardiac surgery. With a successful predictive model, patients will be at a lower risk for premature extubation and prolonged ventilation, lowering ventilation complications and mortality. Overall, the information provided by ML to the physician will allow more optimized and personalized decisions to be made for the patient by incorporating multivariable and multilayer analysis on a patient to patient basis. With smarter decisions being made through time-series analysis of more patient data, the goal of increasing ventilation weaning success rates and preventing prolonged ventilation and complications is within reach.
Prediction of ARDS
Acute respiratory distress syndrome (ARDS) is a life threatening clinical syndrome characterized by acuity of onset, decreased oxygen saturation, capillary endothelial damage, and pulmonary infiltrations. Risk factors for ARDS include advanced age, female gender, smoking, alcohol use, aortic vascular surgery, and cardiovascular surgery.
Similar to ventilation weaning, using ML to predict ARDS can allow the treating physician the opportunity to initiate changes in ventilation strategies to minimize the chance of ventilator associated trauma, work of breathing, and the fibroproliferative phase of ARDS seen in prolonged mechanical ventilation. The pathology of ARDS requires treatment with positive airway pressure and lower tidal volumes, however it can be difficult to optimize a patient to ideal airway volumes. With the use of ML, ventilation parameters like peak airway pressures, resistive pressures, waveform, flow rate, and others can be collected and analyzed in real-time using time-series analysis to optimize and personalize ventilation settings for patients. If the onset of ARDS is caught early enough or even before occurrence, ventilation requiring intubation may be altogether avoidable.
Prediction of Postoperative Atrial Fibrillation and Ventricular Tachycardia
The most common complication following cardiac surgery is atrial fibrillation (AF), occurring upwards of 30% after coronary artery bypass grafting (CABG) and 40% after valve replacements or repairs. AF has many known complications including congestive heart failure (CHF), a three times higher risk of postoperative stroke, and renal insufficiency. Antiarrhythmic prophylactic measurements are frequently utilized in order to lower the rate of POAF, but the American Heart Association (AHA) reports that 60-80% of patients are exposed to unnecessary pharmacological prophylaxis. Ventricular tachyarrhythmias (VTA), while less common, are more life threatening and may lead to cardiac arrest.This overuse of resources, time, money and exposure to medication side effects has to be measured against the risk of developing a life-threatening arrhythmia.
The ability to accurately determine if a patient will enter AF or VTA following cardiac surgery could drastically improve patient outcomes, decrease the number of patients receiving unnecessary antiarrhythmics, and prevent episodes of sudden cardiac death (SCD) following cardiac surgery. Prediction of VTA is often even more difficult, as it is typically diagnosed by ECG at time of onset. This lapse in knowledge highlights the need for accurate and reliable prediction of POAF and VTA’s prior to their occurrence to allow the treating physician time to administer appropriate prophylactic care and to prevent the overuse of unnecessary pharmacologics.
Data incorporation into the ML ecosystem will be the driving force of predictive model development. Figure 1 highlights data pools available to the ML ecosystem in terms of algorithm development, but determining what data to use in the ecosystem remains to be answered.
Physician notes are readily available and can include preoperative risk factors such as sex, weight, valvular status, previous ischemic injury, hypertension, heart failure, or any previous history of arrhythmias. Preoperative, intraoperative, and postoperative physical exam findings such as volume status, pulse rate, blood pressure, and oxygen saturation can also be retrieved using a SQL platform which also streams the waveform data into the secured database for analysis before and after an arrhythmic event. Furthermore, intraoperative variables such as cross-clamp time can be recorded for incorporation of real-time output dependent diagnostic decisions by the physician care team.
One of the most commonly utilized tools for arrhythmia detection is the ECG. ECG and telemetry have the potential to be very useful in ML arrhythmia detection due to availability, real-time data recording, and quantifiable data output. With ML, telemetry data can be recorded and analyzed in real-time, using time-series analysis to predict future arrhythmic events.
The goal of ML in arrhythmia prediction should be to provide clinicians with more information to accurately decide which patients will benefit from prophylactic intervention to prevent POAF and VTA. With a successful predictive model using time-series analysis, this goal is certainly reachable. Taye et al. has set a good foundation for the possibility of physician intervention prior to arrhythmia, and with continued development of AI and machine learning, the time between patient alarm and arrhythmia onset may be increased to allow more time for care team intervention or immediate bedside resuscitation upon arrest.
Prediction of postoperative AKI
Acute kidney injury (AKI) after cardiac surgery is a serious complication due to rapid deterioration in renal function and a marked decrease in glomerular filtration rate (GFR) and renal blood flow (RBF). The pathophysiology surrounding AKI is complex and multifactorial. In cardiac surgery, it is thought that pre-renal causes of decreased RBF and GFR is a significant contributor to AKI, but other factors such as microembolization, neurohormonal activation, exogenous and endogenous toxins, metabolic as well as hemodynamic and inflammation factors, ischemia–reperfusion injury and oxidative stress play important roles as well.30 Even more so, the complicated nature of postoperative AKI diagnosis is contributed to by occult renal insufficiency. Given the complexity surrounding the pathophysiology of AKI onset, ML could be utilized to allow physicians the opportunity to prophylactically prevent AKI and to identify AKI earlier in its pathological course, preventing further damage.
In its current state, ML appears to be very promising for the prediction of AKI following cardiac surgery. Rank et al. has set a foundation for other institutions looking to use ML in AKI prediction through the use of dynamic time-series patient data and by comparing their model to physician accuracy. Given their results, it is possible that ML may replace or be used in combination with some of the older risk models of AKI development such as KDIGO, STS, or THAKAR risk models. The benefit of ML compared to risk models is the personalization of risk stratification by patient rather than lumping all patients into a predefined and limited variable set. ML is able to combine and analyze much larger data sets than risk scores and perform multi-layer analysis on that data set, essentially giving individualized risks for AKI development. As shown in the study done by Rank et al., these models may predict AKI’s better than physicians, allowing earlier treatment and possibly prevention of AKI altogether. Clinically, early identification of AKI can lead to point of care changes that include earlier volume resuscitation, avoidance of nephrotoxic contrast agents prior to or after surgery, or close electrolyte monitoring and replacement.
Conclusion
This paper summarizes current uses of ML in predictive models for ventilation weaning, ARDS, POAF, VTA, and AKI with the goal of understanding how ML can be used in the postoperative CICU. ML augments the physician’s clinical decision making by analyzing massive amounts of high fidelity data to predict disease states based on learning models. With the extensive data analysis performed by ML, a personalized and more comprehensive risk stratification for postoperative disease may be achievable. ML is best used as a tool for the physician, and in combination with clinical context, physical exam, and clinical gestalt, the treating physician will be able to make more informed and individualized care decisions based on the needs of the patient.
The current efforts in ML are promising and may be the next step towards preventative care in medicine.
