Machine Learning and Artificial Intelligence in Neurosurgery: Status, Prospects, and Challenges

Ethical Incorporation of Artificial Intelligence into Neurosurgery: A Generative Pretrained Transformer Chatbot-Based, Human-Modified Approach

INTRODUCTION

Artificial intelligence (AI) has become increasingly used in neurosurgery. Generative pretrained transformers (GPTs) have been of particular interest. However, ethical concerns regarding the incorporation of AI into the field remain underexplored. We delineate key ethical considerations using a novel GPT-based, human-modified approach, synthesize the most common considerations, and present an ethical framework for the involvement of AI in neurosurgery.

METHODS

GPT-4, ChatGPT, Bing Chat/Copilot, You, Perplexity.ai, and Google Bard were queried with the prompt "How can artificial intelligence be ethically incorporated into neurosurgery?". Then, a layered GPT-based thematic analysis was performed. The authors synthesized the results into considerations for the ethical incorporation of AI into neurosurgery. Separate Pareto analyses with 20% threshold and 10% threshold were conducted to determine salient themes. The authors refined these salient themes.

RESULTS

Twelve key ethical considerations focusing on stakeholders, clinical implementation, and governance were identified. Refinement of the Pareto analysis of the top 20% most salient themes in the aggregated GPT outputs yielded 10 key considerations. Additionally, from the top 10% most salient themes, 5 considerations were retrieved. An ethical framework for the use of AI in neurosurgery was developed.

CONCLUSIONS

It is critical to address the ethical considerations associated with the use of AI in neurosurgery. The framework described in this manuscript may facilitate the integration of AI into neurosurgery, benefitting both patients and neurosurgeons alike. We urge neurosurgeons to use AI only for validated purposes and caution against automatic adoption of its outputs without neurosurgeon interpretation.

Beyond human in neurosurgical exams: ChatGPT's success in the Turkish neurosurgical society proficiency board exams

Chat Generative Pre-Trained Transformer (ChatGPT) is a sophisticated natural language model that employs advanced deep learning techniques and is trained on extensive datasets to produce responses akin to human conversation for user inputs. In this study, ChatGPT's success in the Turkish Neurosurgical Society Proficiency Board Exams (TNSPBE) is compared to the actual candidates who took the exam, along with identifying the types of questions it answered incorrectly, assessing the quality of its responses, and evaluating its performance based on the difficulty level of the questions. Scores of all 260 candidates were recalculated according to the exams they took and included questions in those exams for ranking purposes of this study. The average score of the candidates for a total of 523 questions is 62.02 ± 0.61 compared to ChatGPT, which was 78.77. We have concluded that in addition to ChatGPT's higher response rate, there was also a correlation with the increase in clarity regardless of the difficulty level of the questions with Clarity 1.5, 2.0, 2.5, and 3.0. In the participants, however, there is no such increase in parallel with the increase in clarity.

Future of Artificial Intelligence in Surgery: A Narrative Review

Artificial intelligence (AI) is the capability of a machine to execute cognitive processes that are typically considered to be functions of the human brain. It is the study of algorithms that enable machines to reason and perform mental tasks, including problem-solving, object and word recognition, and decision-making. Once considered science fiction, AI today is a fact and an increasingly prevalent subject in both academic and popular literature. It is expected to reshape medicine, benefiting both healthcare professionals and patients. Machine learning (ML) is a subset of AI that allows machines to learn and make predictions by recognizing patterns, thus empowering the medical team to deliver better care to patients through accurate diagnosis and treatment. ML is expanding its footprint in a variety of surgical specialties, including general surgery, ophthalmology, cardiothoracic surgery, and vascular surgery, to name a few. In recent years, we have seen AI make its way into the operating theatres. Though it has not yet been able to replace the surgeon, it has the potential to become a highly valuable surgical tool. Rest assured that the day is not far off when AI shall play a significant intraoperative role, a projection that is currently marred by safety concerns. This review aims to explore the present application of AI in various surgical disciplines and how it benefits both patients and physicians, as well as the current obstacles and limitations facing its seemingly unstoppable rise.

Virtual Reality in Preoperative Planning of Complex Cranial Surgery

OBJECTIVE

Changing paradigms of neurosurgical training and limited operative exposure during the residency period have made it necessary to evaluate newer technologies for training. Virtual reality (VR) technology provides three-dimensional reconstruction of routine imaging, along with the ability to see as well as interact. The application of VR technology in operative planning, which is an important part of neurosurgical training, has been incompletely studied so far.

METHODS

Sixteen final-year residents, post-M.Ch. (magister chirurgiae) residents, and fellows were included as study participants. They were divided into 2 groups based on their seniority for further analysis. Five complex cranial cases were selected and a multiple-choice question-based test was prepared by the authors, with 5 questions for each of the cases. The pretest score was determined based on performance on the test after participants accessed routine preoperative imaging. The posttest score was calculated after use of the VR system (ImmersiveTouch VR System, ImmersiveTouch Inc.). Analysis was performed by the investigators, who were blinded to the identity of the participant. Subanalysis based on the type of case and type of question was performed. Feedback was obtained from each participant regarding VR use.

RESULTS

There was an overall improvement in scores from pretest to posttest, which was also noted in the analysis based on the participants' seniority. This improvement was noted to be more for the vascular cases (15.89%) compared with the tumor cases (7.84%). Participants also fared better in questions related to surgical anatomy and surgical approach, compared with questions based on the diagnosis. There was overall positive feedback from participants regarding VR use, and most participants wanted VR to become a routine part of operative planning.

CONCLUSIONS

Our study shows that there is improvement in understanding of surgical aspects after use of this VR system.

A Pilot Remote Curriculum to Enhance Resident and Medical Student Understanding of Machine Learning in Healthcare

BACKGROUND

Despite the expanding role of machine learning (ML) in health care and patient expectations for clinicians to understand ML-based tools, few for-credit curricula exist specifically for neurosurgical trainees to learn basic principles and implications of ML for medical research and clinical practice. We implemented a novel, remotely delivered curriculum designed to develop literacy in ML for neurosurgical trainees.

METHODS

A 4-week pilot medical elective was designed specifically for trainees to build literacy in basic ML concepts. Qualitative feedback from interested and enrolled students was collected to assess students' and trainees' reactions, learning, and future application of course content.

RESULTS

Despite 15 interested learners, only 3 medical students and 1 neurosurgical resident completed the course. Enrollment included students and trainees from 3 different institutions. All learners who completed the course found the lectures relevant to their future practice as clinicians and researchers and reported improved confidence in applying and understanding published literature applying ML techniques in health care. Barriers to ample enrollment and retention (e.g., balancing clinical responsibilities) were identified.

CONCLUSIONS

This pilot elective demonstrated the interest, value, and feasibility of a remote elective to establish ML literacy; however, feedback to increase accessibility and flexibility of the course encouraged our team to implement changes. Future elective iterations will have a semiannual, 2-week format, splitting lectures more clearly between theory (the method and its value) and application (coding instructions) and will make lectures open-source prerequisites to allow tailoring of student learning to their planned application of these methods in their practice and research.

Computer-vision based analysis of the neurosurgical scene - A systematic review

Introduction

With increasing use of robotic surgical adjuncts, artificial intelligence and augmented reality in neurosurgery, the automated analysis of digital images and videos acquired over various procedures becomes a subject of increased interest. While several computer vision (CV) methods have been developed and implemented for analyzing surgical scenes, few studies have been dedicated to neurosurgery.

Research question

In this work, we present a systematic literature review focusing on CV methodologies specifically applied to the analysis of neurosurgical procedures based on intra-operative images and videos. Additionally, we provide recommendations for the future developments of CV models in neurosurgery.

Material and methods

We conducted a systematic literature search in multiple databases until January 17, 2023, including Web of Science, PubMed, IEEE Xplore, Embase, and SpringerLink.

Results

We identified 17 studies employing CV algorithms on neurosurgical videos/images. The most common applications of CV were tool and neuroanatomical structure detection or characterization, and to a lesser extent, surgical workflow analysis. Convolutional neural networks (CNN) were the most frequently utilized architecture for CV models (65%), demonstrating superior performances in tool detection and segmentation. In particular, mask recurrent-CNN manifested most robust performance outcomes across different modalities.

Discussion and conclusion

Our systematic review demonstrates that CV models have been reported that can effectively detect and differentiate tools, surgical phases, neuroanatomical structures, as well as critical events in complex neurosurgical scenes with accuracies above 95%. Automated tool recognition contributes to objective characterization and assessment of surgical performance, with potential applications in neurosurgical training and intra-operative safety management.

GPT-4 Artificial Intelligence Model Outperforms ChatGPT, Medical Students, and Neurosurgery Residents on Neurosurgery Written Board-Like Questions

BACKGROUND

Artificial intelligence (AI) and machine learning have transformed health care with applications in various specialized fields. Neurosurgery can benefit from artificial intelligence in surgical planning, predicting patient outcomes, and analyzing neuroimaging data. GPT-4, an updated language model with additional training parameters, has exhibited exceptional performance on standardized exams. This study examines GPT-4's competence on neurosurgical board-style questions, comparing its performance with medical students and residents, to explore its potential in medical education and clinical decision-making.

METHODS

GPT-4's performance was examined on 643 Congress of Neurological Surgeons Self-Assessment Neurosurgery Exam (SANS) board-style questions from various neurosurgery subspecialties. Of these, 477 were text-based and 166 contained images. GPT-4 refused to answer 52 questions that contained no text. The remaining 591 questions were inputted into GPT-4, and its performance was evaluated based on first-time responses. Raw scores were analyzed across subspecialties and question types, and then compared to previous findings on Chat Generative pre-trained transformer performance against SANS users, medical students, and neurosurgery residents.

RESULTS

GPT-4 attempted 91.9% of Congress of Neurological Surgeons SANS questions and achieved 76.6% accuracy. The model's accuracy increased to 79.0% for text-only questions. GPT-4 outperformed Chat Generative pre-trained transformer (P < 0.001) and scored highest in pain/peripheral nerve (84%) and lowest in spine (73%) categories. It exceeded the performance of medical students (26.3%), neurosurgery residents (61.5%), and the national average of SANS users (69.3%) across all categories.

CONCLUSIONS

GPT-4 significantly outperformed medical students, neurosurgery residents, and the national average of SANS users. The mode's accuracy suggests potential applications in educational settings and clinical decision-making, enhancing provider efficiency, and improving patient care.

Daily Pain Prediction Using Smartphone Speech Recordings of Patients With Spine Disease

BACKGROUND

Pain evaluation remains largely subjective in neurosurgical practice, but machine learning provides the potential for objective pain assessment tools.

OBJECTIVE

To predict daily pain levels using speech recordings from personal smartphones of a cohort of patients with diagnosed neurological spine disease.

METHODS

Patients with spine disease were enrolled through a general neurosurgical clinic with approval from the institutional ethics committee. At-home pain surveys and speech recordings were administered at regular intervals through the Beiwe smartphone application. Praat audio features were extracted from the speech recordings to be used as input to a K-nearest neighbors (KNN) machine learning model. The pain scores were transformed from a 0 to 10 scale to low and high pain for better discriminative capacity.

RESULTS

A total of 60 patients were enrolled, and 384 observations were used to train and test the prediction model. Using the KNN prediction model, an accuracy of 71% with a positive predictive value of 0.71 was achieved in classifying pain intensity into high and low. The model showed 0.71 precision for high pain and 0.70 precision for low pain. Recall of high pain was 0.74, and recall of low pain was 0.67. The overall F1 score was 0.73.

CONCLUSION

Our study uses a KNN to model the relationship between speech features and pain levels collected from personal smartphones of patients with spine disease. The proposed model is a stepping stone for the development of objective pain assessment in neurosurgery clinical practice.

The role of an open artificial intelligence platform in modern neurosurgical education: a preliminary study

The use of artificial intelligence in neurosurgical education has been growing in recent times. ChatGPT, a free and easily accessible language model, has been gaining popularity as an alternative education method. It is necessary to explore the potential of this program in neurosurgery education and to evaluate its reliability. This study aimed to show the reliability of ChatGPT by asking various questions to the chat engine, how it can contribute to neurosurgery education by preparing case reports or questions, and its contributions when writing academic articles. The results of the study showed that while ChatGPT provided intriguing and interesting responses, it should not be considered a dependable source of information. The absence of citations for scientific queries raises doubts about the credibility of the answers provided. Therefore, it is not advisable to solely rely on ChatGPT as an educational resource. With further updates and more specific prompts, it may be possible to improve its accuracy. In conclusion, while ChatGPT has potential as an educational tool, its reliability needs to be further evaluated and improved before it can be widely adopted in neurosurgical education.

Deep learning characterization of brain tumours with diffusion weighted imaging

Glioblastoma multiforme (GBM) is one of the most deadly forms of cancer. Methods of characterizing these tumours are valuable for improving predictions of their progression and response to treatment. A mathematical model called the proliferation-invasion (PI) model has been used extensively in the literature to model the growth of these tumours, though it relies on known values of two key parameters: the tumour cell diffusivity and proliferation rate. Unfortunately, these parameters are difficult to estimate in a patient-specific manner, making personalized tumour forecasting challenging. In this paper, we develop and apply a deep learning model capable of making accurate estimates of these key GBM-characterizing parameters while simultaneously producing a full prediction of the tumour progression curve. Our method uses two sets of multi sequence MRI in order to produce estimations and relies on a preprocessing pipeline which includes brain tumour segmentation and conversion to tumour cellularity. We first apply our deep learning model to synthetic tumours to showcase the model's capabilities and identify situations where prediction errors are likely to occur. We then apply our model to a clinical dataset consisting of five patients diagnosed with GBM. For all patients, we derive evidence-based estimates for each of the PI model parameters and predictions for the future progression of the tumour, along with estimates of the parameter uncertainties. Our work provides a new, easily generalizable method for the estimation of patient-specific tumour parameters, which can be built upon to aid physicians in designing personalized treatments.

Machine learning in clinical diagnosis, prognostication, and management of acute traumatic spinal cord injury (SCI): A systematic review

Background

Machine learning has been applied to improve diagnosis and prognostication of acute traumatic spinal cord injury. We investigate potential for clinical integration of machine learning in this patient population to navigate variability in injury and recovery.

Materials and methods

We performed a systematic review using PRISMA guidelines through PubMed database to identify studies that use machine learning algorithms for clinical application toward improvements in diagnosis, management, and predictive modeling.

Results

Of the 132 records identified, a total of 13 articles met inclusion criteria and were included in final analysis. Of the 13 articles, 5 focused on diagnostic accuracy and 8 were related to prognostication or management of traumatic spinal cord injury. Across studies, 1983 patients with spinal cord injury were evaluated with most classifying as ASIA C or D. Retrospective designs were used in 10 of 13 studies and 3 were prospective. Studies focused on MRI evaluation and segmentation for diagnostic accuracy and prognostication, investigation of mean arterial pressure in acute care and intraoperative settings, prediction of ambulatory and functional ability, chronic complication prevention, and psychological quality of life assessments. Decision tree, random forests (RF), support vector machines (SVM), hierarchical cluster tree analysis (HCTA), artificial neural networks (ANN), convolutional neural networks (CNN) machine learning subtypes were used.

Conclusions

Machine learning represents a platform technology with clinical application in traumatic spinal cord injury diagnosis, prognostication, management, rehabilitation, and risk prevention of chronic complications and mental illness. SVM models showed improved accuracy when compared to other ML subtypes surveyed. Inherent variability across patients with SCI offers unique opportunity for ML and personalized medicine to drive desired outcomes and assess risks in this patient population.

Machine learning for outcome prediction of neurosurgical aneurysm treatment: Current methods and future directions

INTRODUCTION

Machine learning algorithms have received increased attention in neurosurgical literature for improved accuracy over traditional predictive methods. In this review, the authors sought to assess current applications of machine learning for outcome prediction of neurosurgical treatment of intracranial aneurysms and identify areas for future research.

METHODS

A PRISMA-compliant systematic review of the PubMed, MEDLINE, and EMBASE databases was conducted for all studies utilizing machine learning for outcome prediction of intracranial aneurysm treatment. Patient characteristics, machine learning methods, outcomes of interest, and accuracy metrics were recorded from included studies.

RESULTS

16 studies were ultimately included in qualitative synthesis. Studies primarily analyzed angiographic outcomes, functional outcomes, or complication prediction using clinical, radiological, or composite variables. The majority of included studies utilized supervised learning algorithms for analysis of dichotomized outcomes.

CONCLUSIONS

Commonly included variables were demographics, presentation variables (including ruptured or unruptured status), and treatment used. Areas for future research include increased generalizability across institutions and for smaller datasets, as well as development of front-end tools for clinical applicability of published algorithms.

Machine Learning-Based Surgical Planning for Neurosurgery: Artificial Intelligent Approaches to the Cranium

Objectives

Artificial intelligence (AI) applications in neurosurgery have an increasing momentum as well as the growing number of implementations in the medical literature. In recent years, AI research define a link between neuroscience and AI. It is a connection between knowing and understanding the brain and how to simulate the brain. The machine learning algorithms, as a subset of AI, are able to learn with experiences, perform big data analysis, and fulfill human-like tasks. Intracranial surgical approaches that have been defined, disciplined, and developed in the last century have become more effective with technological developments. We aimed to define individual-safe, intracranial approaches by introducing functional anatomical structures and pathological areas to artificial intelligence.

Methods

Preoperative MR images of patients with deeply located brain tumors were used for planning. Intracranial arteries, veins, and neural tracts are listed and numbered. Voxel values of these selected regions in cranial MR sequences were extracted and labeled. Tumor tissue was segmented as the target. Q-learning algorithm which is a model-free reinforcement learning algorithm was run on labeled voxel values (on optimal paths extracted from the new heuristic-based path planning algorithm), then the algorithm was assigned to list the cortico-tumoral pathways that aim to remove the maximum tumor tissue and in the meantime that functional anatomical tissues will be least affected.

Results

The most suitable cranial entry areas were found with the artificial intelligence algorithm. Cortico-tumoral pathways were revealed using Q-learning from these optimal points.

Conclusions

AI will make a significant contribution to the positive outcomes as its use in both preoperative surgical planning and intraoperative technique equipment assisted neurosurgery, its use increased.

Code-free machine learning for object detection in surgical video: a benchmarking, feasibility, and cost study

OBJECTIVE

While the utilization of machine learning (ML) for data analysis typically requires significant technical expertise, novel platforms can deploy ML methods without requiring the user to have any coding experience (termed AutoML). The potential for these methods to be applied to neurosurgical video and surgical data science is unknown.

METHODS

AutoML, a code-free ML (CFML) system, was used to identify surgical instruments contained within each frame of endoscopic, endonasal intraoperative video obtained from a previously validated internal carotid injury training exercise performed on a high-fidelity cadaver model. Instrument-detection performances using CFML were compared with two state-of-the-art ML models built using the Python coding language on the same intraoperative video data set.

RESULTS

The CFML system successfully ingested surgical video without the use of any code. A total of 31,443 images were used to develop this model; 27,223 images were uploaded for training, 2292 images for validation, and 1928 images for testing. The mean average precision on the test set across all instruments was 0.708. The CFML model outperformed two standard object detection networks, RetinaNet and YOLOv3, which had mean average precisions of 0.669 and 0.527, respectively, in analyzing the same data set. Significant advantages to the CFML system included ease of use, relatively low cost, displays of true/false positives and negatives in a user-friendly interface, and the ability to deploy models for further analysis with ease. Significant drawbacks of the CFML model included an inability to view the structure of the trained model, an inability to update the ML model once trained with new examples, and the inability for robust downstream analysis of model performance and error modes.

CONCLUSIONS

This first report describes the baseline performance of CFML in an object detection task using a publicly available surgical video data set as a test bed. Compared with standard, code-based object detection networks, CFML exceeded performance standards. This finding is encouraging for surgeon-scientists seeking to perform object detection tasks to answer clinical questions, perform quality improvement, and develop novel research ideas. The limited interpretability and customization of CFML models remain ongoing challenges. With the further development of code-free platforms, CFML will become increasingly important across biomedical research. Using CFML, surgeons without significant coding experience can perform exploratory ML analyses rapidly and efficiently.

The National Inpatient Sample: A Primer for Neurosurgical Big Data Research and Systematic Review

OBJECTIVE

The National Inpatient Sample - the largest all-payer inpatient database in the United States - is an important instrument for big data analysis of neurosurgical inquiries. However, earlier research has determined that many NIS studies are limited by common methodological pitfalls. In this study, we provide the first primer of NIS methodological procedures in the setting of neurosurgical research and review all published neurosurgical studies utilizing the NIS.

METHODS

We designed a protocol for neurosurgical big data research using the NIS, based on the authors' subject matter expertise, NIS documentation, and input and verification from the Healthcare Cost and Utilization Project. We subsequently used a comprehensive search strategy to identify all neurosurgical studies utilizing the NIS in the PubMed and MEDLINE, Embase, and Web of Science databases from inception to August 2021. Studies underwent qualitative categorization (years of the NIS studied, neurosurgical subspecialty, age group, and thematic focus of study objective) and analysis of longitudinal trends.

RESULTS

We identified a canonical, four-step protocol for NIS analysis: study population selection, defining additional clinical variables, identification and coding of outcomes, and statistical analysis. Methodological nuances discussed include identifying neurosurgery-specific admissions, addressing missing data, calculating additional severity and hospital-specific metrics, coding perioperative complications, and applying survey weights to make nationwide estimates. Inherent database limitations and common pitfalls of NIS studies discussed include lack of disease process-specific variables and data following the index admission, inability to calculate certain hospital-specific variables after 2011, performing state-level analyses, conflating hospitalization charges and costs, and not following proper statistical methodology for performing survey-weighted regression. In a systematic review, we identified 647 neurosurgical studies utilizing the NIS. While almost 60% of studies were published after 2015, <10% of studies analyzed NIS data after 2015. The average sample size of studies was 507,352 patients (standard deviation=2,739,900). Most studies analyzed cranial procedures (58.1%) and adults (68.1%). The most prevalent topic areas analyzed were surgical outcome trends (35.7%) and health policy and economics (17.8%), while patient disparities (9.4%) and surgeon or hospital volume (6.6%) were the least studied.

CONCLUSIONS

We present a standardized methodology to analyze the NIS, systematically review the state of the NIS neurosurgical literature, suggest potential future directions for neurosurgical big data inquiries, and outline recommendations to improve the design of future neurosurgical data instruments.

Artificial neural networks predict the need for permanent cerebrospinal fluid diversion following posterior fossa tumor resection

Background

Resection of posterior fossa tumors (PFTs) can result in hydrocephalus that requires permanent cerebrospinal fluid (CSF) diversion. Our goal was to prospectively validate a machine-learning model to predict postoperative hydrocephalus after PFT surgery requiring permanent CSF diversion.

Methods

We collected preoperative and postoperative variables on 518 patients that underwent PFT surgery at our center in a retrospective fashion to train several statistical classifiers to predict the need for permanent CSF diversion as a binary class. A total of 62 classifiers relevant to our data structure were surveyed, including regression models, decision trees, Bayesian models, and multilayer perceptron artificial neural networks (ANN). Models were trained using the (N = 518) retrospective data using 10-fold cross-validation to obtain accuracy metrics. Given the low incidence of our positive outcome (12%), we used the positive predictive value along with the area under the receiver operating characteristic curve (AUC) to compare models. The best performing model was then prospectively validated on a set of 90 patients.

Results

Twelve percent of patients required permanent CSF diversion after PFT surgery. Of the trained models, 8 classifiers had an AUC greater than 0.5 on prospective testing. ANNs demonstrated the highest AUC of 0.902 with a positive predictive value of 83.3%. Despite comparable AUC, the remaining classifiers had a true positive rate below 35% (compared to ANN, P < .0001). The negative predictive value of the ANN model was 98.8%.

Conclusions

ANN-based models can reliably predict the need for ventriculoperitoneal shunt after PFT surgery.