Predictive model for preoperative risk calculation of cerebrospinal fluid leak after resection of midline craniofacial mass lesions.

Background: Complex anterior skull base defects produced by resection of mass lesions vary in size and configuration and may be extensive. We analyzed the largest single-center series of midline craniofacial lesions extending intra- and extracranially. The study aims at the development of a predictive model for preoperative measurement of the risk of the postoperative cerebrospinal fluid (CSF) leak based on patients' characteristics and surgical plans. Methods: 166 male and 149 female patients with mean age 40,5 years (1 year and - 81 years) operated for benign and tumor-like midline craniofacial mass lesions were retrospectively analyzed using logistic regression method (Ridge regression algorithm was selected). The overall CSF leak rate was 9.6%. The ROSE algorithm and 'glmnet' software suite in R were used to overcome the cohort's disbalance and avoid overtraining the model. Results: The most influential modifiable negative predictor of the postoperative CSF leak was the use of extracranial and combined approaches. Use of transbasal approaches, gross total resection, utilization of one or two vascularized flaps for skull base reconstruction were the foremost modifiable predictors of a good outcome. Criterium of elevated risk was established at 50% with a specificity of the model as high as 0.83. Conclusions: The performed study has allowed for identifying the most significant predictors of postoperative CSF leak and developing an effective formula to estimate the risk of this complication using data known for each patient. We believe that the suggested web-based online calculator can be helpful for decision making support in off-pattern clinical situations.

Predicting the Length of Stay in Neurosurgery with RuGPT-3 Language Model.

In this study, we update the evaluation of the Russian GPT3 model presented in our previous paper in predicting the length of stay (LOS) in neurosurgery. We aimed to assess the performance the Russian GPT-3 (ruGPT-3) language model in LOS prediction using narrative medical records in neurosurgery compared to doctors' and patients' expectations. Doctors appeared to have the most realistic LOS expectations (MAE = 2.54), while the model's predictions (MAE = 3.53) were closest to the patients' (MAE = 3.47) but inferior to them (p = 0.011). A detailed analysis showed a solid quality of ruGPT-3 performance based on narrative clinical texts. Considering our previous findings obtained with recurrent neural networks and FastText vector representation, we estimate the new result as important but probably improveable.

Noninvasive Glioma Grading with Deep Learning: A Pilot Study.

Gliomas are the most common neuroepithelial brain tumors, different by various biological tissue types and prognosis. They could be graded with four levels according to the 2007 WHO classification. The emergence of non-invasive histological and molecular diagnostics for nervous system neoplasms can revolutionize the efficacy and safety of medical care and radically reduce healthcare costs. Our pilot study aimed to evaluate the diagnostic accuracy of deep learning (DL) in subtyping gliomas by WHO grades (I-IV) based on preoperative magnetic resonance imaging (MRI) from Burdenko Neurosurgery Center's database. A total of 707 MRI studies was included. A "3D classification" approach predicting tumor type for the entire patient's MRI data showed the best result (accuracy = 83%, ROC AUC = 0.95), consistent with that of other authors who used different methodologies. Our preliminary results proved the separability of MR T1 axial images with contrast enhancement by WHO grade using DL.

A Comparison of Word Embeddings to Study Complications in Neurosurgery.

Our study aimed to compare the capability of different word embeddings to capture the semantic similarity of clinical concepts related to complications in neurosurgery at the level of medical experts. Eighty-four sets of word embeddings (based on Word2vec, GloVe, FastText, PMI, and BERT algorithms) were benchmarked in a clustering task. FastText model showed the best close to the medical expertise capability to group medical terms by their meaning (adjusted Rand index = 0.682). Word embedding models can accurately reflect clinical concepts' semantic and linguistic similarities, promising their robust usage in medical domain-specific NLP tasks.

Semiautomatic Identification of Pulmonary Embolism in Electronic Health Records Through Sentence Labeling.

In this study, we tested the quality of the information extraction algorithm proposed by our group to detect pulmonary embolism (PE) in medical cases through sentence labeling. Having shown a comparable result (F1 = 0.921) to the best machine learning method (random forest, F1 = 0.937), our approach proved not to miss the information of interest. Scoping the number of texts under review down to distinct sentences and introducing labeling rules contributes to the efficiency and quality of information extraction by experts and makes the challenging tasks of labeling large textual datasets solvable.

Length of Stay Prediction in Neurosurgery with Russian GPT-3 Language Model Compared to Human Expectations.

Patients, relatives, doctors, and healthcare providers anticipate the evidence-based length of stay (LOS) prediction in neurosurgery. This study aimed to assess the quality of LOS prediction with the GPT3 language model upon the narrative medical records in neurosurgery comparing to doctors' and patients' expectations. We found no significant difference (p = 0.109) between doctors', patients', and model's predictions with neurosurgeons tending to be more accurate in prognosis. The modern neural network language models demonstrate feasibility in LOS prediction.

Inter-Rater Reliability of Unstructured Text Labeling: Artificially vs. Naturally Intelligent Approaches.

Unstructured medical text labeling technologies are expected to be highly demanded since the interest in artificial intelligence and natural language processing arises in the medical domain. Our study aimed to assess the agreement between experts who judged on the fact of pulmonary embolism (PE) in neurosurgical cases retrospectively based on electronic health records and assess the utility of the machine learning approach to automate this process. We observed a moderate agreement between 3 independent raters on PE detection (Light's kappa = 0.568, p = 0). Labeling sentences with the method we proposed earlier might improve the machine learning results (accuracy = 0.97, ROC AUC = 0.98) even in those cases that could not be agreed between 3 independent raters. Medical text labeling techniques might be more efficient when strict rules and semi-automated approaches are implemented. Machine learning might be a good option for unstructured text labeling when the reliability of textual data is properly addressed. This project was supported by the RFBR grant 18-29-22085.

Semiautomated Approach for Muscle Weakness Detection in Clinical Texts.

The automated detection of adverse events in medical records might be a cost-effective solution for patient safety management or pharmacovigilance. Our group proposed an information extraction algorithm (IEA) for detecting adverse events in neurosurgery using documents written in a natural rich-in-morphology language. In this paper, we challenge to optimize and evaluate its performance for the detection of any extremity muscle weakness in clinical texts. Our algorithm shows the accuracy of 0.96 and ROC AUC = 0.96 and might be easily implemented in other medical domains.

Detection of Muscle Weakness in Medical Texts Using Natural Language Processing.

Identifying adverse events in clinical documents is demanded in retrospective clinical research and prospective monitoring of treatment safety and cost-effectiveness. We proposed and evaluated a few methods of semi-automated muscle weakness detection in preoperative clinical notes for a larger project on predicting paresis by images. The combination of semi-expert and machine learning methods demonstrated maximized sensitivity = 0.860 and specificity = 0.919, and largest AUC = 0.943 with a 95% CI [0.874; 0.991], outperforming each method used individually. Our approaches are expected to be effective for autoshaping a well- verified training dataset for supervised machine learning.