Almost one in five patients benefitted from a preoperative ultrasound before undergoing bariatric metabolic surgery (BMS) with 15.9% of patients found to have chronic calculous cholecystitis and an additional 1.4% patients having their BMS postponed due to other findings, according to a study led by researchers from Alexandria University, Alexandria, Egypt. However, although these findings indicate a potential benefit in certain cases, the study’s authors caution that it is necessary to fully evaluate routine preoperative ultrasound’s overall utility and economic impact (cost–benefit analysis) in BMS patients.
The study’s findings, reported in the paper, ‘The Role of Preoperative Abdominal Ultrasound in the Preparation of Patients Undergoing Primary Metabolic and Bariatric Surgery: A Machine Learning Algorithm on 4418 Patients’ Records’, published in Obesity Surgery, also explored how machine learning (ML) models could significantly enhance healthcare professional’s ability to make informed decisions and improving patient outcomes through more precise and predictive analytics.
The utilisation of preoperative abdominal ultrasonography (US) in evaluating patients with obesity before BMS is contentious, with some surgeons arguing that it can detect intra-abdominal anomalies (eg. gallbladder and biliary tract issues) or organ enlargement that could influence outcomes during or post-surgery. Others however see it as unnecessary as it is time-consuming, not cost-effective, rarely alters the surgical approach and its effectiveness in patients with obesity may be limited due to excessive soft tissue.
The study’s authors routinely performed abdominal US on all patients undergoing BMS and performed a retrospective analysis of patients who underwent primary BMS. d to assess the outcomes of routine preoperative US and explore the role of alternative radiological techniques in cases necessitating further diagnostic evaluation. Additionally, machine learning techniques were applied to identify any variables with significant predictive capabilities, aiding in developing a clinical prediction model that effectively identifies patients likely to benefit from a preoperative US examination.
Patients were categorised into the following four groups based on ultrasound results:
Group 1 consisted of patients with normal ultrasound results.
Group 2 consisted of patients with non-significant findings that did not impact the planned procedure.
Group 3 consisted of patients with findings that did not affect the surgical plan but required concomitant surgery and/or postoperative follow-up.
Group 4 consisted of patients with significant findings that directly affected the procedure or required further radiological, laboratory, or endoscopic investigations. Group 4 was then divided into two subgroups:
Group 4A included patients with findings that did not impact the surgical plan but delayed the surgery until other radiological investigations were completed.
Group 4B included patients whose findings directly affected the surgical plan, resulting in either postponement for assessment by another specialty or cancellation of the procedure.
Outcomes
In total, the medical records of 5,720 patients were analysed and after applying the exclusion criteria, 4,418 patients’ records were included in the study. The majority of patients were female (70.7%), the average age was 43.3±13.8 years and the mean BMI was 48.1±7.5kg/m2.
Group 1 was the largest at 45.7%. Group 2 comprised 35.7% and had minor, non-impactful findings. Group 3, making up 17.0%, required additional surgery or follow-up without altering the original surgical plan. Group 4, the smallest at 1.5%, included significant findings that affected the procedure; this included Group 4A (0.8%), where additional imaging caused delays, and Group 4B (0.7%), where surgeries were either cancelled (0.3%) or postponed (0.4%).
In Group 2, the most prevalent finding was fatty liver and hepatomegaly, accounting for 87.7% across the group and 31.28% of the total cohort. In Group 3, chronic calculous cholecystitis was the most prevalent condition, representing 93.2% and 15.89% of the total cohort.
In Group 4, the most prevalent conditions were hepatic focal lesions at 20.3% within the group (0.32% from the total cohort), renal lesions at 13.0% (0.20% from the total cohort), and pancreatic lesions at 10.1% (0.16% from the total cohort).
In Group 4A, findings did not impact the surgical plan but delayed the surgery until other radiological investigations were completed. Group 4B directly affected the surgical plan, resulting in either postponement or cancellation of the procedure.
In total, 19 cases (0.4% total cohort, 57.6% of this group) were postponed, with a range of final diagnoses varied from several cysts, presents of stones, suspected Crohn’s disease, sarcoidosis or serous cystadenoma.
In 14 cases (0.3% of the total cohort, 42.4% of this group), BMS was cancelled due to the range of final diagnoses varied from oesophageal leiomyoma to non-Hodgkin lymphoma, several liver metastases, and tumours (ovarian, pancreatic, intraductal papillary-mucinous and lymphoma).
Concomitant surgeries included chronic calculous cholecystitis (n=702, 93.2%), gallbladder polyps (n=7, 0.9%), ovarian dermoid cysts (n=2, 0.3%) and inguinal hernias (n=7, 0.9%), which were found in Group 3.
Machine Learning
The ML model used in this study identified several key predictive variables: BMI, diabetes status, smoking habits, HCV previous infection, WBC count, FT4 levels, platelet count, AST activity and TSH levels.
“Correlations based on data that are “invisible to the naked eye,” whereby such insights could advance our understanding by identifying underlying variables that are predictors in workup processes, guiding when the US would be a logical next step in specific clinical presentations,” the authors write. “However, external validation of these results is necessary to determine if these characteristics are consistent across other studies.”
The researchers argued that by leveraging large datasets, ML can uncover complex patterns and relationships that may not be apparent through traditional statistical methods, therefore offering more profound insights into disease mechanisms and patient responses. Indeed, they stated that ML could revolutionise predictive healthcare by enabling more precise and earlier interventions, ultimately reducing costs and improving outcomes.
However, issues around data privacy, integration ML with real-time data from electronic health records, medical ethics, non-maleficence, justice and autonomy, to name a few, need to be addressed to ensure ML applications in healthcare are effective, equitable, and trustworthy, aligning with the principles of ethical medical practice.
The authors noted that ultrasound imaging is highly operator-dependent, with significant variabilities in training and interpretation among technicians and radiologists leading to inconsistencies in detecting and classifying conditions. In addition, different ultrasound machines and settings also produce variations in image quality. Therefore, they called for future studies to standardise ultrasound protocols and provide extensive operator training to mitigate these limitations.
“In summary, ultrasound’s operator-dependent nature and variability in imaging techniques are limitations that can affect study outcomes,” they stated. “Acknowledging these factors underscores the need for standardisation and training to enhance the reliability of ultrasound assessments.”
To access this paper, please click here
Comments