Keywords
AI - CSF - machine learning - NLP - skull base - BERT - reconstruction - rhinology
Introduction
The endoscopic endonasal approach has become the dominant and widely accepted method
for removing skull base tumors in the sellar and parasellar regions.[1] Skull base reconstruction plays a pivotal role in mitigating the risk of cerebrospinal
fluid (CSF) leakage and its associated complications, including the development of
meningitis.[2] Achieving successful skull base reconstruction and preventing CSF leaks has significant
positive implications for reducing hospitalization duration and alleviating the burden
on healthcare costs.[3] CSF leak is a recognized risk during endoscopic sellar surgery, especially when
managing large tumors or craniopharyngiomas. Evidence indicates that intraoperative
CSF leaks occur in about 30.1% of cases, while postoperative leaks are reported in
roughly 3.7 to 9%,[4]
[5] influenced by various factors such as surgeon experience, tumor characteristics,
and individual patient-related variables. The risk of CSF leak following endoscopic
skull base surgery is further affected by tumor size and consistency, patient age,
BMI, and surgical factors[6]; prevention depends on appropriate reconstruction techniques, with autologous grafts
used for small, low-flow defects and vascularized tissue, particularly the nasoseptal
flap, serving as the gold standard for larger or high-flow leaks in anterior skull
base reconstruction.[7]
Despite established reconstruction techniques and known risk factors, accurately predicting
which patients will develop postoperative CSF leaks remains challenging, necessitating
more sophisticated predictive methodologies.
The advancement of artificial intelligence in healthcare enables the development of
machine learning (ML) models that can predict CSF leak risk factors.[7] AI models can integrate multiple patient, tumor, and surgical variables to predict
intraoperative CSF leaks. In a cohort of 238 pituitary adenoma cases, machine-learning
methods were tested; random forest performed best, accurately identifying most leaks
and outperforming traditional statistics and other ML models.[8]
This capability will significantly impact surgical practice by informing pre-, intra-,
and postoperative management decisions, including reconstruction method selection.
In natural language processing (NLP), BERT has become a benchmark model because it
is first exposed to vast amounts of raw, unlabeled text and can process information
from both directions across relatively long passages. BERT, short for bidirectional
encoder representations from transformers, uses only the encoder part of the transformer
architecture to learn word meanings that depend on the words that come before and
after them, producing a highly adaptable, context-aware language representation. Although
the original BERT was trained on general-purpose English text, researchers have since
built tailored versions for specialized domains. For example, BioBERT, BlueBERT, SciBERT,
and PubMedBERT have been retrained on biomedical literature so the model recognizes
technical vocabulary and clinical phrasing more accurately. These domain-specific
variants preserve BERT's key strengths, bidirectional context, and the ability to
digest long sequences while improving performance on biomedical text-mining tasks.[9] This is the first study to compare a BERT-based model with conventional statistical
techniques for detecting CSF-leak risk factors in endoscopic skull-base surgery, leveraging
qualitative (nonnumeric) inputs to simplify the handling of large, variable datasets
and ultimately boost predictive accuracy and guide reconstruction decisions more effectively.
Materials and Methods
King Saud University Medical City, in collaboration with the College of Computer Science
at King Saud University. In this study, we developed a traditional logistic regression
model and compared its performance to an NLP model, specifically a BERT model. Ethical
approval for the study was obtained from the ethics committee at the College of Medicine
at King Saud University (no. 23-8232). [Fig. 1] shows the overall methodological process of this project.
Fig. 1 Methodological overview of the study process.
Data Categorization
We extracted information from healthcare records within the study setting, including
all cases of skull base pathologies managed by a multidisciplinary team specializing
in rhinology, skull base surgery, and neurosurgery at King Saud University Medical
City. The available data, which served as input for both the BERT model and the traditional
statistical model, were categorized into the following groups: demographics: age,
gender, and BMI. Perioperative clinical CSF leak indicators: postoperative clinical
symptoms or signs raising suspicion, need for postoperative CT scans and their timing,
and length of hospital stay. Pathology-related factors: tumor size, presence of hydrocephalus,
tumor location, and histopathology. Surgical factors: type and extent of surgical
approach, preoperative CSF diversion, intraoperative CSF leak presence, CSF leak flow
rate, nasal flap reconstruction, use of nasal packing, and types of reconstruction
materials used. Perioperative CT scan features, including the timing, presence, and
location of pneumocephalus; the presence and size of any hematoma; graft characteristics;
signs of air communication; and findings from repeat CT scans, were evaluated by a
skull-base team comprising a rhinologist, neurosurgeon, and radiologist. The resulting
radiological findings data were then entered manually into a text-based datasheet.
Development of the Traditional Logistic Regression Model
Descriptive statistics were generated using the compareGroups package, with missing
data imputed (Hmisc), categorical variables factorized, and class imbalance addressed
via upsampling (caret). Continuous data were summarized as mean ± SD or median/IQR,
and categorical data as counts and percentages. Univariate analysis used Somers' D
and the C statistic (AUC) to assess variable discrimination. Clinically relevant variables
with high C statistics were included regardless of significance. Upsampling outperformed
SMOTE and ROSE, preserving data structure and improving model performance. A bootstrap
approach reduced overfitting, and nonsignificant variables were excluded. The final
logistic regression model, validated by ROC and kappa statistics, showed improved
predictive accuracy and robustness.
Development of the BERT Model
Categorical medical data input: previously described were collected as input data,
forming the foundation of the predictive model.
Tabular Dataset
Each patient record is represented as a sequence x = [c1,…,c⅞,y], where c contains k categorical attributes encoded as their text labels (e.g., “male,” “Craniopharyngioma”)
and y∈{0,1} denotes the presence of a CSF leak. The entire text string is passed to the standard
BERT tokenizer (bert-base-uncased), which automatically converts the words into WordPiece
tokens, adds the special [CLS] and [SEP] markers, and pads or truncates the sequence to a fixed length.
The resulting token IDs are fed to the BERT encoder. Within each layer, multi-head
self-attention computes contextualized representations using the canonical (see [Supplementary Material S1], available in the online version only, for formulas details), followed by residual
connections, layer normalization, and feed-forward sublayers. After the final layer,
the [CLS] token is prepended to each sequence, and its final hidden state is extracted
as a pooled representation and passed through a linear layer to classify CSF leak
versus no leak, as illustrated in [Fig. 2].
Fig. 2 Overview of the classification pipeline based on BERT for CSF leak detection.
All mathematical derivations, hyperparameters, and the training loss are presented
in [Supplementary Material S1] (available in the online version only).
BERT Model Training and Optimization
The BERT classifier was fine-tuned on twofold cross-validation (50% training/50% testing).
Optimization used AdamW (learning-rate 1 × 10−5, weight-decay 1 × 10−2) for 20 epochs and a binary cross-entropy objective; focal weighting was applied
to mitigate class imbalance (see [Supplementary Material S2], available in the online version only, for the exact loss formula). Performance
was reported as the mean of the two folds using accuracy, precision, recall (sensitivity),
positive and negative predictive values, F1 score (the harmonic mean of precision
and recall, ranging from 0 to 1), and AUC.
Results
Descriptive Statistics for the Study Subjects
A total of 116 skull-base surgery patients were included. No statistically significant
differences were found between the CSF-leak and no-leak groups in age (median: 29
years vs. 42 years, p = 0.526), gender (60.0% vs. 57.4% female, p = 1.000), preoperative CSF diversion (53.3% vs. 50.5%, p = 1.000), or hydrocephalus (26.7% vs. 8.9%, p = 0.065). Weight, tumor location, pathology size, reconstruction materials (fat,
fascia, hard graft, flap, nasal pack, glue, gasket), and surgical approach also showed
no significant differences. Moreover, within both the gross-total resection (GTR)
and subtotal resection (STR) groups, the proportions of CSF-leak versus no-leak cases
were not significantly different (all p > 0.05) as detailed in [Table 1].
Table 1
Descriptive statistics for the study sample
|
No CSF leak
|
CSF leak
|
p-Value
|
|
n = 101
|
n = 15
|
|
|
Age
|
42.0 [24.0; 54.0]
|
29.0 [23.5; 49.0]
|
0.526
|
|
Gender
|
|
|
1.000
|
|
Female
|
58 (57.4%)
|
9 (60.0%)
|
|
|
Male
|
43 (42.6%)
|
6 (40.0%)
|
|
|
Preoperative diversion
|
|
|
1.000
|
|
No
|
51 (50.5%)
|
7 (46.7%)
|
|
|
Yes
|
50 (49.5%)
|
8 (53.3%)
|
|
|
Preoperative hydrocephalus
|
|
|
0.065
|
|
No
|
92 (91.1%)
|
11 (73.3%)
|
|
|
Yes
|
9 (8.91%)
|
4 (26.7%)
|
|
|
Weight
|
|
|
|
|
Underweight
|
5 (4.95%)
|
0 (0.00%)
|
1.000
|
|
Normal
|
31 (30.7%)
|
5 (33.3%)
|
1.000
|
|
Overweight
|
21 (20.8%)
|
6 (40.0%)
|
0.112
|
|
Obese
|
44 (43.6%)
|
4 (26.7%)
|
0.338
|
|
Location (anterior skull base)
|
|
|
0.580
|
|
No
|
52 (51.5%)
|
6 (40.0%)
|
|
|
Yes
|
49 (48.5%)
|
9 (60.0%)
|
|
|
Location (posterior fossa)
|
|
|
0.633
|
|
No
|
92 (91.1%)
|
13 (86.7%)
|
|
|
Yes
|
9 (8.91%)
|
2 (13.3%)
|
|
|
Location (sellar/suprasellar)
|
|
|
0.413
|
|
No
|
59 (58.4%)
|
11 (73.3%)
|
|
|
Yes
|
42 (41.6%)
|
4 (26.7%)
|
|
|
Size of the pathology
|
|
|
|
|
(1–2) cm
|
16 (15.8%)
|
3 (20.0%)
|
0.710
|
|
(2–3) cm
|
28 (27.7%)
|
4 (26.7%)
|
1.000
|
|
>3 cm
|
44 (43.6%)
|
8 (53.3%)
|
0.666
|
|
Type of reconstruction material
|
|
|
|
|
Fat
|
77 (76.2%)
|
13 (86.7%)
|
0.515
|
|
Fascia
|
34 (33.7%)
|
6 (40.0%)
|
0.849
|
|
Hard (bone)
|
42 (41.6%)
|
10 (66.7%)
|
0.122
|
|
Naso-septal flap
|
65 (64.4%)
|
11 (73.3%)
|
0.695
|
|
nasal pack
|
74 (73.3%)
|
9 (60.0%)
|
0.358
|
|
Glue
|
67 (66.3%)
|
10 (66.7%)
|
1.000
|
|
Gasket
|
21 (20.8%)
|
3 (20.0%)
|
1.000
|
|
Endoscopic approach
|
|
|
0.238
|
|
Extended approach
|
61 (60.4%)
|
12 (80.0%)
|
|
|
Standard approach (transsphenoidal transsellar)
|
40 (39.6%)
|
3 (20.0%)
|
|
|
GTR (gross total resection)
|
|
|
0.532
|
|
No
|
53 (52.5%)
|
6 (40.0%)
|
|
|
Yes
|
48 (47.5%)
|
9 (60.0%)
|
|
|
STR (subtotal resection)
|
|
|
1.000
|
|
No
|
64 (63.4%)
|
10 (66.7%)
|
|
|
Yes
|
37 (36.6%)
|
5 (33.3%)
|
|
Performance and Results of Logistic Regression Model
Odds Ratio of Different Variables
Preoperative hydrocephalus (odds ratio [OR] = 5.15, p = 0.016), overweight status (OR = 7.15, p < 0.001), fat use in reconstruction (OR = 19.65, p < 0.001), and gross total resection (OR = 3.83, p = 0.001) were all associated with increased odds of CSF leak. In contrast, nasal
pack (OR = 0.03, p < 0.001), gasket use (OR = 0.20, p = 0.003), and sellar/suprasellar tumor location (OR = 0.10, p < 0.001) significantly reduced leak risk as shown in [Table 2].
Table 2
Logistic regression analysis results
|
Dependent variable: CSF leak
|
Odds ratios
|
Confidence interval
|
p-Value
|
|
(Intercept)
|
0.11
|
0.03–0.40
|
0.001
|
|
Preoperative hydrocephalus
|
5.15
|
1.46–21.61
|
0.016
|
|
Overweight
|
7.15
|
2.98–18.60
|
<0.001
|
|
Nasal pack
|
0.03
|
0.01–0.10
|
<0.001
|
|
Gasket
|
0.20
|
0.07–0.56
|
0.003
|
|
Fat
|
19.65
|
5.52–82.96
|
<0.001
|
|
Tumor location (sellar/suprasellar vs. other)
|
0.10
|
0.03–0.29
|
<0.001
|
|
Gross total resection (GTR vs. other)
|
3.83
|
1.77–8.79
|
0.001
|
|
Observations
|
202
|
|
|
|
R
2 Tjur
|
0.401
|
|
|
Intraoperative CSF leaks occurred in 20% of the leak group versus 16.8% of the nonleak
group (p = 0.721). Among leak cases, high-flow leaks were more common (13.3% vs. 5.95%), while
low-flow leaks were less frequent (6.67% vs. 14.3%) compared to nonleak cases (p = 0.367). Reconstruction methods varied between groups but showed no significant
difference (p = 0.167). Postoperative CT revealed significantly more and higher-grade pneumocephalus
in the leak group (p = 0.004). Excessive fat was more frequent (58.3% vs. 12.7%, p = 0.001), and fat size was more often inadequate (p = 0.048) in the leak group. Hematoma requiring evacuation was slightly more common
but not significant (p = 0.492). Postoperative CT revealed factors associated with CSF leaks, including
greater extra-cavitary fat graft displacement (p = 0.002) and less favorable solid reconstruction positioning (p = 0.047). Flap adherence was reduced in the leak group (p = 0.060), and air continuity between the nasal and surgical cavities was observed
in only one leak case (p = 0.153). Although resection extent varied: standard approach: transsphenoidal transsellar
versus Extended approach, the difference was not statistically significant (p = 0.862) as shown in [Table 3].
Table 3
Intraoperative and postoperative findings
|
No leak
|
Leak
|
p-Value overall
|
n
|
|
n = 101
|
n = 15
|
|
|
|
Presence of Intraoperative CSF leak
|
|
|
0.721
|
116
|
|
No
|
84 (83.2%)
|
12 (80.0%)
|
|
|
|
Yes
|
17 (16.8%)
|
3 (20.0%)
|
|
|
|
If yes, the CSF flow was
|
|
|
0.367
|
99
|
|
High flow
|
|
2 (13.3%)
|
|
|
|
Low flow
|
|
1 (6.67%)
|
|
|
|
None
|
(100%)
|
12 (80.0%)
|
|
|
|
Type of solid reconstruction
|
|
|
0.167
|
52
|
|
Bone
|
5 (11.9%)
|
3 (30.0%)
|
|
|
|
Medpore
|
34 (81.0%)
|
6 (60.0%)
|
|
|
|
Mesh titanium
|
1 (2.38%)
|
0 (0.00%)
|
|
|
|
Omnipore plate
|
0 (0.00%)
|
1 (10.0%)
|
|
|
|
Plastic plate
|
2 (4.76%)
|
0 (0.00%)
|
|
|
|
Presence of pneumocephalus in the postoperative CT scan
|
|
|
0.004
|
86
|
|
Grade 0 (none)
|
42 (57.5%)
|
2 (15.4%)
|
|
|
|
Grade 1 (dots [<1 mm of air])
|
3 (4.11%)
|
0 (0.00%)
|
|
|
|
Grade 2 (bubbles [<1 cm of air])
|
9 (12.3%)
|
1 (7.69%)
|
|
|
|
Grade 3 (1–3 cm air)
|
4 (5.48%)
|
4 (30.8%)
|
|
|
|
Grade 4 (>3 cm of air)
|
15 (20.5%)
|
6 (46.2%)
|
|
|
|
Presence of a big hematoma: a need for evacuation?
|
|
|
0.492
|
85
|
|
No
|
69 (95.8%)
|
12 (92.3%)
|
|
|
|
Yes
|
3 (4.17%)
|
1 (7.69%)
|
|
|
|
Presence of excessive fat
|
|
|
0.001
|
83
|
|
No
|
62 (87.3%)
|
5 (41.7%)
|
|
|
|
Yes
|
9 (12.7%)
|
7 (58.3%)
|
|
|
|
Size Fat-graft?
|
|
|
0.048
|
82
|
|
All fat outside the surgical cavity
|
0 (0.00%)
|
1 (7.69%)
|
|
|
|
Part of the fat is outside
|
3 (4.35%)
|
2 (15.4%)
|
|
|
|
Good size
|
66 (95.7%)
|
10 (76.9%)
|
|
|
|
The location of fat graft?
|
|
|
0.002
|
77
|
|
started to go outside the surgical cavity
|
9 (13.8%)
|
4 (33.3%)
|
|
|
|
Within the sinus
|
0 (0.00%)
|
2 (16.7%)
|
|
|
|
Within the surgical cavity
|
56 (86.2%)
|
6 (50.0%)
|
|
|
|
Sign of air continuity between the surgical cavity and the nasal cavity
|
|
|
0.153
|
85
|
|
No
|
72 (100%)
|
12 (92.3%)
|
|
|
|
Yes
|
0 (0.00%)
|
1 (7.69%)
|
|
|
|
Solid reconstruction, is it in a good location?
|
|
|
0.047
|
64
|
|
No
|
20 (37.7%)
|
8 (72.7%)
|
|
|
|
Yes
|
33 (62.3%)
|
3 (27.3%)
|
|
|
|
Is the location of the septal flap adherent around all bone defects?
|
|
|
0.060
|
73
|
|
No
|
14 (22.6%)
|
6 (54.5%)
|
|
|
|
Yes
|
48 (77.4%)
|
5 (45.5%)
|
|
|
Reconstruction materials, including fat, fascia lata, hard grafts, nasoseptal flap,
nasal packing, and bone, varied between CSF leak and nonleak groups, but none showed
statistically significant differences (all p > 0.05), as shown in [Table 1]. Histopathological tumor types showed no significant association with CSF leak incidence,
with pituitary adenoma being the most common in leak cases (26.7%), followed by meningioma
(20.0%), chordoma and craniopharyngioma (13.3% each), and all comparisons yielded
nonsignificant p-values as shown in [Table 4].
Table 4
Histopathology type and CSF leak
|
Histopathology diagnosis
|
No leak
|
Leak
|
p-Value overall
|
|
n = 101
|
n = 15
|
|
|
Atypical cartilaginous neoplasm
|
1 (0.99%)
|
0 (0.00%)
|
1.000
|
|
Chordoma
|
7 (6.93%)
|
2 (13.3%)
|
0.328
|
|
Craniopharyngioma
|
6 (5.94%)
|
2 (13.3%)
|
0.276
|
|
Dermoid cyst
|
1 (0.99%)
|
0 (0.00%)
|
1.000
|
|
Epidermoid cyst
|
0 (0.00%)
|
1 (6.67%)
|
0.129
|
|
Fibrous dysplasia
|
2 (1.98%)
|
0 (0.00%)
|
1.000
|
|
Fungal sinusitis
|
4 (3.96%)
|
0 (0.00%)
|
1.000
|
|
Germ cell tumor
|
1 (0.99%)
|
0 (0.00%)
|
1.000
|
|
Juvenile nasopharyngeal angiofibroma
|
3 (2.97%)
|
0 (0.00%)
|
1.000
|
|
Meningioma
|
24 (23.8%)
|
3 (20.0%)
|
1.000
|
|
Meningoencephalocele
|
3 (2.97%)
|
1 (6.67%)
|
0.430
|
|
Pilocytic astrocytoma
|
2 (1.98%)
|
1 (6.67%)
|
0.342
|
|
Pituitary Adenoma
|
43 (42.6%)
|
4 (26.7%)
|
0.374
|
|
Rathke
|
1 (0.99%)
|
0 (0.00%)
|
1.000
|
|
Sarcoma
|
0 (0.00%)
|
1 (6.67%)
|
0.129
|
|
Spontaneous CSF leak
|
2 (1.98%)
|
0 (0.00%)
|
1.000
|
|
Temporal bone fracture
|
1 (0.99%)
|
0 (0.00%)
|
1.000
|
Logistic Regression Model Performance
The logistic regression model achieved an AUC of 0.847, with high specificity (90.6%)
but low sensitivity (21.4%) and a Kappa of 0.1447, indicating slight agreement. McNemar's
test (p < 2e-16) showed significant prediction bias; positive predictive value (PPV) was
44.7%, negative predictive value (NPV) 76.5%, with CSF leak prevalence at 26.2%, detection
rate 5.6%, and detection prevalence 12.6%. The balanced accuracy of 56.0% reflects
poor performance in detecting true CSF leaks due to class imbalance.
Performance and Results of the BERT Model
In terms of performance metrics for the BERT model at the confusion matrix (50:2),
it was found to have an AUC of 1.0000 and an accuracy of 0.9833, a PPV of 0.8889,
an NPV of 1.0000, a recall of 0.9808, and an F1 score of 0.9657
The average grouped attention weights across twofold of the proposed model's training
revealed the following findings: perioperative CT scans received the highest attention
weight by far, followed by surgical factors. The other three groups: demographics,
suspicion of CSF leak, and pathology-related factors (including histopathology type)
had significantly lower attention weights, as shown in [Fig. 3].
Fig. 3 Categorical groups predictors for CSF leak.
Comparison of the Logistic Regression Model versus the BERT Model
The BERT model surpassed the traditional logistic regression model in predicting CSF
leaks, achieving an accuracy of 98%, a PPV of 89%, an NPV of 100%, a recall of 98%,
and an F1 score of 96%. In contrast, the logistic regression model had an accuracy
of 72%, sensitivity of 20%, specificity of 90%, detection rate of 5%, and PPV of 44%,
as illustrated in [Table 5] and [Fig. 4].
Table 5
Comparison of the logistic regression model versus the BERT model
|
Metric
|
Logistic regression model
|
BERT model
|
|
AUC
|
0.847
|
1.0000
|
|
Accuracy
|
0.725
|
0.9833
|
|
95% CI
|
(0.7189, 0.731)
|
|
|
Sensitivity
|
0.2143
|
1.0000
|
|
Specificity
|
0.9060
|
0.9808
|
|
Positive predictive value
|
0.4471
|
0.8889
|
|
Negative predictive value
|
0.7649
|
1.0000
|
|
F1 score
|
0.2897
|
0.9657
|
Fig. 4 Comparison of performance of the BERT model versus the logistic regression model
in the prediction of CSF leak.
Discussion
In this study, we developed a BERT-based NLP model that transforms qualitative medical
data into numerical form for machine learning analysis to predict CSF leaks after
skull base surgery. The model significantly outperformed traditional logistic regression,
achieving 98% accuracy, 89% precision, 98% recall, and a 96% F1 score, compared to
logistic regression's 72% accuracy, 20% sensitivity, 90% specificity, 5% detection
rate, and 44% PPV, highlighting the potential of NLP-based models for superior clinical
outcome prediction.
Our BERT model analysis revealed that perioperative CT scan features suggestive of
CSF leakage carried the highest attention weight for prediction, followed by surgical
factors. These findings align with a large multicenter study that evaluated three
predictive models (logistic regression, decision tree, and neural network) for CSF
leak prediction. Their research identified intraoperative CSF leakage as the most
significant risk factor for CSF rhinorrhea, with elevated BMI and revision surgery
also contributing significantly in transsphenoidal approaches.[10] The consistency between their conclusions and our results reinforces the superior
performance of machine learning methodologies compared to traditional statistical
approaches in predicting this critical surgical complication. The BERT model processed
the full dataset without manual feature selection, revealing that perioperative CT
findings and intraoperative factors were stronger predictors of CSF leak than preoperative
variables. Including early postoperative CT data allowed the model to identify key
indicators of CSF leak at the time of clinical suspicion, highlighting the value of
focusing on perioperative and surgical data for more precise prediction models. A
neural network study for predicting intraoperative CSF leaks during pituitary surgery
demonstrated 88% classification accuracy (AUC: 0.84), outperforming conventional statistical
methods, which identified no significant risk factors. The neural network achieved
high sensitivity (83%) and specificity (89%), with high suprasellar Hardy grade, prior
surgery, and older age as the primary predictive factors.[11] A systematic review of seven AI studies for CSF leak prediction in pituitary surgery
found performance metrics ranging from 0.73 to 0.98 (AUC) and 0.70 to 0.97 (accuracy).
Random Forest was the most frequently used algorithm, with k-fold cross-validation
as the predominant validation method. Notably, deep learning models demonstrated significantly
higher pooled sensitivity than machine learning models (99% vs. 86.2%, p < 0.01), while specificity remained comparable between approaches (90.6% vs. 92.1%,
p = 0.87).[12] Our 12% CSF leak rate aligns with previous studies; elevated BMI was a significant
risk factor (OR = 7.15), consistent with Ivan et al,[13] while Fraser et al,[14] specifically identified BMI > 25 kg/m2 and posterior fossa tumors as predictors of higher postoperative leak rates in their
615-patient EEA study. While our data showed no significant correlation between histopathological
types or tumor location and CSF leak risk, Zhang et al,[15] found clival tumors associated with higher leak rates in their 100 extra-sellar
tumor cases. Our results identified preoperative hydrocephalus as a significant predictor
(OR: 5.15, CI: 1.46–21.61, p = 0.016), consistent with Patel et al,[16] who found only BMI and hydrocephalus as significant predictors in their 806-case
analysis. Our analysis found no statistical significance for various reconstruction
techniques, including solid reconstruction. A multicenter analysis of 706 patients
reported a 7.8% postoperative CSF leak rate, indicating that rigid reconstruction
and older age were protective factors against postoperative sellar leaks, whereas
BMI was not linked to increased risk[17]—contrary to our findings. Kuan et al,[18] reported that only intraoperative CSF leak was associated with recurrence in their
300 consecutive repair cases. A large systematic review and meta-analysis of risk
factors for postoperative CSF leakage after endonasal endoscopic skull-base surgery
reported that overweight and obesity were associated with an OR of 1.88 (95% CI, 1.35
– 2.63; p < 0.01), a result that aligns with our findings. Regarding reconstruction, 16 studies
(total n = 3,579) assessed pedicled vascularized flaps, showing a pooled OR of 0.62 for CSF
leakage compared with free grafts.[19] In contrast, our findings demonstrated an unexpectedly high OR of 19.56 for CSF
leak when fat grafts were used. This finding diverges from prior literature, in a
systematic review of fat graft use in transsphenoidal surgery, postoperative CSF leak
requiring intervention was significantly lower in the fat-in-sphenoid-sinus repair
group (4.4 %) than in multilayer (20.3 %) or no-repair groups (12.6 %; p < 0.01).[20] One plausible explanation for this discrepancy is that fat grafts in our series
were preferentially applied to high-flow defects, which inherently carry a greater
leak risk.
Our study introduces a novel application of an NLP model for CSF leak prediction in
skull base surgery. We believe that the BERT model holds considerable promise for
advancing both clinical practice and research. Its implementation in future studies
with larger sample sizes may lead to improved clinical decision-making, as the model
has demonstrated the ability to identify patterns and associations that traditional
statistical methods may overlook. Furthermore, the application of this model in research
settings could offer significant advantages in handling large and complex datasets,
particularly those involving qualitative variables. Unlike conventional approaches,
BERT's key advantage lies in its ability to directly process qualitative data without
numerical conversion, preserving contextual information while streamlining analysis.
Despite promising results, the exceptionally high-performance metrics suggest potential
overfitting or data leakage. As a single-center study, our findings face generalizability
limitations. Future validation should include a critical review of CT features, rigorous
verification of train/test methodology, validation on independent datasets, and optimization
of model complexity. We recommend multicenter implementation of BERT for CSF leak
analysis to increase sample size and potentially identify risk factors undetectable
through conventional statistical approaches.
Conclusion
BERT NLP model outperforms traditional logistic regression in predicting CSF leaks
after endoscopic skull base surgery, demonstrating superior accuracy with qualitative
clinical data, enhancing risk stratification and decision-making.
