"This study aims to collect data on patients with small bowel obstruction (SBO) admitted to hospitals in France and Italy from May 2022 to October 2024 to develop a deep convolutional neural network (DCNN) model. This model will analyze anonymized CT scans to assess the effectiveness of non-operative management (NOM) for SBO, supporting decisions on surgical intervention. Eligible patients are those diagnosed with SBO due to abdominal adhesions who initially received NOM for at least 24 hours. Patients with other SBO causes, early surgery within 24 hours, or those without a CT scan diagnosis are excluded. Data collection spans hospitals in Antibes, Nice, Milan, and Vimercate, targeting consecutive SBO cases with adhesive etiology. To perform an external validation of the DCNN, data will also be retrospectively collected from patients admitted to the Antibes hospital between May 2021 and April 2022 with adhesive SBO. This validation set includes patients who underwent NOM successfully and those who needed surgery after NOM failure. The DCNN model will be applied to anonymized, non-contrast and contrast-enhanced portal-phase CT scans of these patients, with researchers blinded to each patient's NOM outcome to prevent bias. The model's performance will then be evaluated using accuracy metrics consistent with those used in initial model testing, ensuring the reliability of results when applied to external cases. NOM, after adhesive SBO diagnosis via clinical exams, blood tests, and CT scans, includes fasting, analgesics, antiemetics, and fluids as per current guidelines, without necessarily using nasogastric tubes or contrast agents. Patients are re-evaluated after 24 hours to determine whether NOM should continue or if surgery is necessary. NOM is deemed effective if patients experience symptom resolution, stool passage, and no recurrence within 90 days. NOM failure is defined by the need for laparoscopic or laparotomic surgery, based on symptoms' persistence, worsening, or radiological indicators of blockage despite adequate NOM. Data collection, registered with the French National Committee for Data Protection, includes variables like age, sex, medical history, symptoms, blood tests, CT-scan findings, NOM details, and surgical information. Radiological data, including Digital Imaging and Communication in Medicine (DICOM) files of CT scans, will be anonymized and converted to the Neuroimaging Informatics Technology Initiative (NIfTI) format for secure storage and analysis. The NIfTI data files will be randomly split into training and test datasets in an 80%-20% ratio, processed separately for non-contrast and contrast-enhanced CT scans. Data augmentation, including random rotation, flipping, zooming, translation, and noise addition, will be applied to improve model accuracy and reduce overfitting. Different DCNN models will be trained and tested and furtherly undergo external validation to produce a tool capable of predicting NOM failure and need for surgery in patients with adhesive SBO."
Study Type
OBSERVATIONAL
Enrollment
370
Andrea CHIERICI
Nice, France
Performance of Deep Learning model in predicting NOM failure
To measure the degree of agreement between the prediction made by the Deep Learning model and the actual success or failure of non-operative management for small bowel obstruction in the real situation. This will be measured through Area Under the Receiver Operator Characteristic curve (AUROC) assesment. AUROC varies between 0.5 and 1, corresponding to no class separation capacity and full class separation capacity, respectively.
Time frame: 90 days from patients hospital discharge.
Performance evaluation: accuracy
The number of true positives and true negatives among all predictions. It varies between 0 (no correct prediction) to 1 (full correct predictions).
Time frame: 90 days from patients hospital discharge.
Performance evaluation: precision
The number of true positives divided by all the positive predictions (true positives and false positives). It varies between 0 (no correct prediction) to 1 (full correct predictions)
Time frame: 90 days from patients hospital discharge.
Performance evaluation: recall
The number of true positives divided by the actual positive instances in the dataset (true positives and false negatives). It varies between 0 (no correct prediction) to 1 (full correct predictions).
Time frame: 90 days from patients hospital discharge.
Performance evaluation : balanced accuracy
the aritmethic mean of sensitivity and specificity. It varies between 0 (no correct prediction) to 1 (full correct predictions).
Time frame: 90 days from patients hospital discharge.
Performance evaluation:F1-score
It combines precision and recall. It ranges from 0-100%, and a higher F1 score denotes a better quality classifier.
Time frame: 90 days from patients hospital discharge.
Performance evaluation: Confusion matrix
A visual representation of true positives, false positives, true negatives, and false negatives. It is depicted through a table.
Time frame: 90 days from patients hospital discharge.
Performance evaluation :Log-loss
It indicates how close the prediction probability is to the corresponding actual/true value (0 or 1 in case of binary classification). The more the predicted probability diverges from the actual value, the higher is the log-loss value.
Time frame: 90 days from patients hospital discharge.
Performance evaluation: Cohen's Kappa
A metric used to measure the level of agreement between two raters which can be a useful tool to gauge the performance of a classification model. It accounts for the fact that the raters may happen to agree on some items purely by chance. It varies between 0 (no correct prediction) to 1 (full correct predictions).
Time frame: 90 days from patients hospital discharge.
This platform is for informational purposes only and does not constitute medical advice. Always consult a qualified healthcare professional.