Machine learning in neurosurgery: a global survey

Recent years have witnessed the rise of machine learning applications in the scientific literature, both in basic science and clinical medicine [18, 26]. Neurosurgical practice has always relied on the individual experience of surgeons to carefully balance surgical indications, operative risk and expected outcome [30]. The advent of evidence-based medicine has framed the surgical decision-making process into guidelines based on the results of high-quality data, and of randomized controlled clinical trials—not devoid of several flaws in design themselves [19]. This approach, despite remaining the gold standard, is limited by the oversimplification of patients’ individual characteristics that often do not allow patient-specific analytics. With the exponential growth of data in the era of big data, it is increasingly important to provide clinicians with tools for integrating this individual patient data into reliable prediction models. The latter primarily aims to enhance the surgical decision-making processes and potentially improve outcomes, but predictive analytics furthermore harbour the potential to reduce unnecessary health-care costs [21, 29, 31, 34, 36, 37, 41].

It is often difficult for clinicians to integrate the many described risk factors and outcome predictors into a single workable prognosis [3]. Neurosurgical research and clinical practice is ideal for the application of machine learning (ML), which harbours the potential for predictive analytics to integrate all relevant patient factors in a way that is often too complex for natural intelligence [28, 40]. Moreover, ML can be used to extract deep features from data such as radiological and histological images, or genomic data [16, 38‐40, 43].

At present, the neurosurgical literature is increasingly focusing on substituting traditional statistical models with more complex ML models with the aim of improving predictive power [29, 31]. For example, ML has been used in neurosurgery to predict post-operative satisfaction [2], early post-operative complications [41] or cerebrospinal fluid leaks [37]. Despite this encouraging trend and the presence of recent publications reviewing the large range of publications on ML in neurosurgery [28‐30], data on the worldwide adoption and perception of ML in our specialty are currently lacking. Our aim was to carry out a worldwide survey among neurosurgeons to assess the adoption of ML algorithms into neurosurgical clinical practice and research and to identify factors associated with their use.

There exists no prior published data on the worldwide adoption of ML in neurosurgery. This global survey reached a diverse cohort of neurosurgeons at different levels of training. Our results indicate that ML has already quickly gained wide acceptance in the neurosurgical community, without notable heterogeneity in its global distribution. Almost a third of neurosurgeons reported having applied ML in either clinical practice or research, a value that exceeded expectations. Furthermore, the most common applications of ML in neurosurgery were for prediction of complications and outcomes, as well as to interpret or automatically quantify imaging. No predictors of clinical ML use were identified, again stressing that the availability and acceptance of readily developed ML tools are not bound by socio-demographic factors. On the other hand, among research-active neurosurgeons, some subspecialties as well as academic surgeons appear to apply ML more frequently for their research.

Our study is the first to our knowledge to provide a worldwide overview of the implementation of ML in neurosurgical clinical practice and research. To our surprise, almost a third of respondents stated making use of ML, and this was true for both clinical practice and research. Although this can be partially explained by response bias—with academic surgeons active in the EANS and CNS targeted and with a likely higher response rate to our survey among surgeons interested in ML—our results still indicate that ML is quickly becoming one of the foremost technologies in neurosurgical practice. Importantly, the heterogeneity in adoption rates among regions was relatively low, and adoption of ML into clinical practice was not apparently influenced by limitations in costs or socioeconomic status, as is the case with other less accessible technologies such as robotics [33, 35]. While the development of ML models can often be expensive and resource-intensive, the application of readily trained ML algorithms does not usually require especially high technological standards or expenses. Many ML applications are web-based [25]. For this reason, we expect that ML will increasingly enable enhanced diagnostic, prognostic and predictive analytics around the world, even in the most rural areas.

After controlling for potential confounding factors, we could not identify factors associated with increased or decreased use of ML in clinical practice. This again demonstrates how homogenously ML use seems to be distributed among the neurosurgical community. On the other hand, subspecialists in neuro-oncology, functional neurosurgery, trauma and epilepsy were significantly more likely to apply ML in their research. As expected, surgeons working in non-academic centres and private practice were less likely to engage in ML-based neurosurgical applications, consistent with the development of ML models currently being rather confined to academic institutions possessing the resources, protected time, expertise, extensive databases and computational power to create and distribute algorithms. However, it has to be considered that the development of e.g. ML-based prediction models has been massively eased by free software packages released by the major technology companies, which nowadays enable training of simple ML models on even the most basic notebooks. Still, the development of models may be limited by a lack of high-quality, structured datasets [24].

In fact, ML has already been broadly applied to several subspecialties in neurosurgery spanning from cranial [1, 7, 39], vascular [15, 32], spinal [5, 11, 13, 25, 31, 36] and radiosurgery, among others [23, 41]. Several examples of how ML outperforms traditional statistics and prognostic indexes commonly applied in the clinical practice are already available in the medical literature. For example, a recent study by van Niftrik et al. reported the use of a gradient boosting machine to predict early post-operative complications after intracranial tumour surgery [41]. The authors were able to show improved performance with respect to conventional statistical modelling based on logistic regression and interestingly observed that among the variables in their model, features that were not taken into account in the statistical model, such as histology, anatomical localization or surgical access in fact contributed strongly in the ML model [41]. Oermann et al. also showed that artificial neural networks performed better at 1-year survival prediction than more traditional models in patients with brain metastases treated with radiosurgery [22]. The same group also was able to show an improvement in predictions of arteriovenous malformation radiosurgery outcomes [23]. Staartjes et al. found that a deep learning approach was significantly better at predicting intraoperative cerebrospinal fluid leaks and gross total resection in pituitary surgery than logistic regression, while no predictors could be identified using traditional interferential statistics for the former outcome [34, 37].

In spinal neurosurgery, applications of ML have included prediction of outcome in patients with lumbar disc herniation and lumbar spinal stenosis [2, 31, 36], or to predict complications following elective adult spinal deformity procedures [14]. For example, Khor et al. developed a prediction model from a state-wide database to predict clinically relevant improvement after lumbar spinal fusion and integrated their model into a freely available web app, which was then externally validated [13, 25].. Again, this shows that while it may be resource-intensive to develop such models, they can be rolled out to clinicians and patients around the world for free using simple interfaces.

Radiological applications are ideally suited to machine learning algorithms given the magnitude and complexity of data extractable from examinations such as CT and MRI scans. Interestingly, ML models can establish a hidden relationship between deep radiological features (“radiomics”) and outcomes of the pathology of interest. Lao et al., for example, were able to stratify patients into different prognostic subgroups based on radiomic features [17]. Similarly, it has been shown that it is possible to identify IDH mutation status in gliomas from radiomic features alone [4]. Finally, more extravagant applications of ML in neuroradiology include e.g. the generation of synthetic CT images—practically indistinguishable from actual CTs—from cranial MRI [6, 42].

Despite these positive results, still many present and future potential ML applications remain unknown to the majority of neurosurgical specialists. Our study determined that the factors deterring the use of ML were, in decreasing order, lack of skilled resources (staff, equipment) to develop a model, time limitations restricting ML application in clinical practice, lack of ML models for the indications of interest, uncertainty concerning which processes may benefit most from the application of ML algorithms, as well as—less importantly—lack of data to develop a model, and lack of personal convincement of the added value of this new technology.

Our results warrant some considerations. First, once a ML model with clinical relevance is developed and after it has been externally validated [25], the focus has to shift on making it easy to implement and widely available in clinical practice. Web-based apps that are clinician- or patient-friendly are ideal [12, 13, 25]. Second, while a large proportion of neurosurgeons may already be applying ML in their clinical practice, it is important to foster ML literacy in the neurosurgical community. As with randomized studies forming the basis of evidence-based practice, clinicians should be able to make an informed decision as to which ML models published are likely valid and have applied good methodology, and which ones should probably not be trusted in clinical practice. Lastly, ML relies on the availability of “big data” to be exploited for algorithm training and validation subsequently [21, 24]. A wide and complete collection of patient data in the sense of population-based databases enables more representative ML models. Integrated databases with automated comprehensive data collection that are necessary for such applications are currently few and far between, preventing the development of highly generalizable models [20, 21, 24, 27].

Survey-based studies, while able to provide important insights, have inherent limits because of several potential biases. During survey distribution, selection and response bias are frequent. Time constraints on responders may have limited their ability to answer with maximal accuracy, and in fact concerning the adoption of ML into clinical research, we obtained several incomplete or blank answers. The data is mostly based on subjective impressions of surgeons. Knowing this, bias could arise from the fact that surgeons who are more exposed to neurosurgical ML can value it more positively than those who do not routinely make use of it, and vice-versa. However, the reasons for advantages and disadvantages were specifically captured separately for users and non-users. Additionally, the relative percentage of geographic regions was skewed in favour of western countries, limiting the sensitivity of our survey for what concerns regions such as Asia and Pacific, South America and in particular Africa with only two responses.

This study provides a first global overview of the adoption of ML into neurosurgical practice. Machine learning has the potential to improve diagnostic work-up and neurosurgical decision-making by shedding light on radiological interpretation, surgical outcome and complication prediction and as a consequence patients’ quality of life and surgical satisfaction. A relevant proportion of neurosurgeons appears to already have adopted ML into their clinical practice in some form. The homogenous distribution of ML users in neurosurgery is a testimony to the accessibility of readily developed ML algorithms, even in low-resource settings. Still, many structural issues need to be addressed in order for ML to achieve its full potential in neurosurgery. These include easy-to-access resources for surgeons and patients; prospective-integrated data collection systems to allow model development; and surgeon education on ML, all of which can add to the rapid development of ML in neurosurgery while ensuring high quality of the introduced tools and their correct application. Best practice recommendations, external validation and sound methodology are necessary for any ML tool before its application in our high-stakes clinical practice. Furthermore, future trials may be conducted to assess the real clinical impact—and any changes in decision-making—that may be caused by ML algorithms in neurosurgery.