Artificial intelligence in obstetric anaesthesia

Abstract

‘AI is not going to replace physicians, but physicians who use AI are going to replace physicians who don’t, and that may be the cautionary tale.’ Keith Horvath, MD, former National Institutes of Health director, Cardiothoracic Surgery Research Program. The advent of artificial intelligence is revolutionising the way we work. The essence of artificial intelligence is intelligence exhibited by machines, notably computer systems. Obstetric anaesthesia, in contrast, relies on the synergy of natural and emotional intelligence of living beings. Artificial intelligence is widely anticipated to change how we live, affecting all aspects of society, including healthcare. We already have some insights into how artificial intelligence may work in clinical medicine; tasks that traditionally relied on clinicians, such as x-ray interpretation, are now augmented by automated processes.1 The principles of artificial intelligence are complex, which can make it challenging to fully understand and appreciate. This can potentially lead to clinicians being discouraged from using artificial intelligence technologies. We aim to give an overview of the key concepts relevant to anaesthetists and outline how artificial intelligence may influence obstetric anaesthesia over the next decade. Artificial intelligence is the umbrella term used in computing for machine learning, natural language processing, neural networks, and computer vision.2 At its most basic level, machine learning is an algorithm that performs a task that it has not been specifically programmed or trained to do, enabling computers to learn from data and make decisions or predictions ‘independently’. It ranges from supervised learning trained on the interpretation of large data sets to reinforced learning whereby the algorithm attempts a task and will finetune itself to improve its performance such as autopilot total intravenous anaesthesia (Fig. 1).Fig. 1: Overview of the mechanisms of machine learning.The original language of computer software was in the form of code, such as JAVA, a coding language widely used for developing software applications across various platforms, where a specific input would yield an outcome. Natural language processing is a branch of artificial intelligence designed to make computers understand, interpret, and generate human language in a way that reflects how people communicate, often in real time. It can not only understand words, but interprets the arrangement of phrases to estimate the human meaning. Language translation software is a commonly used form of natural language processing in healthcare. Neural networks are a programme mapped after the human brain, which, like neurons, can make complex nonlinear connections (Fig. 2). Deep learning is an advanced self-learning feature of neural networks that can efficiently analyse very large data sets. A web search engine is built of complex neural networks.Fig. 2: Neural networks.The use of artificial intelligence in anaesthesia remains in its infancy. Early studies across multiple fields show promising results and give an insight into how artificial intelligence may affect clinical practice.3 Risk prediction is an area that has been studied extensively, particularly in the setting of peri-operative hypotension.4 Few studies to date have examined the role of hypotension prediction in the parturient. Gratz et al.5 demonstrated in a pilot study that neural network analysis of continuous noninvasive blood pressure monitoring of arterial stiffness at elective caesarean section accurately predicts hypotension. Primitive models that do not incorporate machine learning, such as closed-loop feedback systems, whereby a pump delivers a bolus of phenylephrine when a blood pressure threshold is met have failed to show convincing results.6,7 The next step that warrants investigation is incorporating hypotension predictive algorithms with a novel automated pump. A pump that treats hypotension before it becomes clinically apparent while anticipating a patient's response, to reduce the risk of over or under-correction, may ultimately improve patient-centred outcomes such as nausea and vomiting. Neuraxial analgesia and anaesthesia are the foundations of obstetric anaesthesia. As obesity becomes more widely seen in the obstetric caseload, strategies to improve the success of neuraxial techniques become increasingly important.8,9 A major area of interest is using artificial intelligence to support ultrasound assisted neuraxial techniques. A machine learning approach was evaluated by Chan et al.10 to identify the needle insertion site for spinal anaesthesia in a cohort of patients with obesity. A successful first attempt at dural-puncture was achieved in 79.1% of cases with a strong correlation between posterior complex depth identification and clinician loss of resistance. In addition, artificial intelligence interpretation of anticoagulation guideline recommendations for neuraxial anaesthesia has shown promising results but showed a reliance on accurate and relevant prompts.11 Artificial intelligence informed decision-making and assistance with neuraxial procedures has the potential to improve subarachnoid block success and reduce neuraxial complications. The role of machine learning in predicting neuraxial anaesthetic dosing has been studied in pregnant women. Wei et al.12 created a decision-support model by designing an equation of regression, a subtype of supervised machine learning. They studied the influence of abdominal girth and vertebral column height in determining neuraxial dose to achieve a block height of T4-T6. The algorithm achieved a positive result in 75% of cases. This was a small study in a homogenous patient cohort, which examined only two factors that may contribute to block height. It does, however, show promise for future larger studies that incorporate a myriad of components, including surgical factors, patient factors and addition of an opioid, for example, to optimise neuraxial dosing. Labour is a complex and multifaceted element of a woman's birth experience and is acknowledged as being one of the most painful experiences a woman will undergo. Epidural analgesia has been the recognised gold standard for intrapartum pain relief for some time.13 Despite this, variations in protocols exist, which usually include patient-controlled epidural analgesia with either a continuous epidural infusion or a more efficacious programmed intermittent epidural bolus.14 The first computer-integrated patient controlled epidural analgesia (CIPCEA) showed encouraging results by adjusting the basal infusion based on the woman's analgesia requirements in the preceding hour.15 Tan et al.16 failed to predict breakthrough pain using a machine learning risk-prediction model. This study included a limited number of static variables such as maternal age, BMI and preneuraxial oxytocin infusion rate. Machine learning analysis of dynamic intrapartum variables such as analgesia requirements and physiological variable monitoring have yet to be studied and potentially provide a strategy to optimise labour analgesia. Natural language processing is already widely accessible through open artificial intelligence chatbots such as ChatGPT. A recent study examined the role artificial intelligence chatbots play in providing unregulated patient information regarding labour analgesia options.17 It highlighted the potential dangers of misinformation and even the possibility of an increased health disparity for pregnant women through inherent biases that exist in artificial intelligence programming. Natural language processing has also been used in the peri-operative setting to extract data from electronic health records including prepopulated options on electronic forms but also the advanced ability to analyse free text clinical notes. It has been shown to outperform anaesthetists by identifying information relevant to an anaesthetic.18 How information is generated and accessed will be influenced by artificial intelligence. It is probable that automated information gathering will reduce the administrative burden and ultimately may enhance patient safety in a speciality with a high emergency workload that is often out-of-hours.8 There is also a potential for artificial intelligence to use this type of information gathering and generate artificial intelligence clinical trials.19 This would be of particular importance to pregnant patients who often find themselves excluded from standard clinical trials because of concerns about potential harm to the foetus or impact on the pregnancy.20 Using artificial intelligence to enable greater participation of pregnant women in clinical trials has the potential to increase access to well tolerated and effective treatments for pregnant women and reduce inequality in outcomes. The influence of artificial intelligence in obstetrics will also affect anaesthetic practice. Cardiotocography (CTG) is the gold standard intrapartum monitoring tool for detecting foetal hypoxia through foetal heart rate variability and maternal uterine contractions. It is widely used for continuous assessment of foetal wellbeing but has a sensitivity of 40 to 60% with high rates of misinterpretation and inter-person variability.21 Despite this, it often directs decisions around emergency caesarean section. The most extensive trial of a computerised alert system to date was the INFANT study, involving more than 46 000 pregnant women across 24 hospitals in the UK and Ireland.22 It aimed to evaluate the impact of the adjunctive clinical decision support provided by the INFANT system on neonatal and maternal outcomes compared to standard care without the system. The study found no difference in the neonatal or maternal outcomes between the two groups. However, artificial intelligence assisted interpretation of foetal heart rate patterns has shown more promise. Fergus et al.23 reported that machine learning and neural network analysis of CTG interpretation is more reliable than obstetricians and midwives with a sensitivity and specificity of over 90%. Early identification of women who may require an operative delivery would allow for better planning of anaesthesia care and potentially improve the overall intrapartum multidisciplinary collaboration and communication. In debates about technology, the role of emotional intelligence is often overlooked. Studies examining the empathy of artificial intelligence have been surprisingly positive.24 The obstetric anaesthetist has an unparalleled role in identifying and alleviating patient concerns which are often subtle and nonverbal. Their role also involves recognising and appreciating patient concerns during a life-changing event, something that machine learning will probably never replace. Although artificial intelligence may have some role in helping to improve empathy it is hard to envisage how it could ever replace the unique intricacies of human interaction and that aspect of the role of an obstetric anaesthetist. The concept of artificial intelligence and how it can affect our practice may be somewhat abstract at present. Anaesthetists have an augmented curiosity, and will no doubt question the introduction of black box algorithms without clarity as to how certain predictions are made. The introduction of the unpublished bispectral index (BIS) algorithm as a depth of anaesthesia monitor raised many concerns about its reliability, but despite that it has now been widely adopted into routine practice.25 Anaesthetic interventions guided by artificial intelligence are being investigated at an accelerating pace. The evidence for its use and trials to assess its application in obstetric anaesthesia to date remain scarce and warrant further investigation. Caution must be taken when extracting data from trials that do not include pregnant women as algorithms may not be translatable to obstetric practice. We must ensure that ethical standards and clinical governance of artificial intelligence meet the same benchmarks as other areas of clinical research.26 As we begin incorporating artificial intelligence into clinical practice, legal and moral dilemmas must be addressed with transparent algorithms that limit black-box predictions to ensure integration into obstetric anaesthesia has the right emphasis on patient-centred care. Despite the growing interest in artificial intelligence, many challenges remain. The infrastructure that exists in many developed healthcare systems is already insufficient to meet its current needs. Technology in healthcare is often replaced and upgraded on a piece-by-piece basis, resulting in an incongruent and fragmented system. Furthermore, education and training in artificial intelligence methodology is currently nonexistent and will probably need to be incorporated into both undergraduate and postgraduate medical curricula.27 Technological framework constraints will slow and even limit the integration of artificial intelligence into healthcare. Despite its challenges, artificial intelligence offers numerous benefits. It can increase efficiency by automating administrative processes and reduce costs by improving care co-ordination, for example, by optimising resources, such as labour ward rooms and staff, based on predicted patient demand and acuity levels. These factors could improve care by enabling clinicians to concentrate on delivering actual patient care. Artificial intelligence heralds an exciting era in technological development and may represent the next big age for mankind. Whilst it is unlikely to replace the role of the obstetric anaesthetist, its presence will no doubt be felt on labour wards and in obstetric theatres in the future. How we embrace artificial intelligence and machine learning for the future remains to be seen.

Ähnliche Arbeiten