2018
December 22, 2023Centroid®
December 29, 2023Article of the Month – December 2023
Development and Validation of Machine Learning Models for Predicting Occult Nodal Metastasis in Early-Stage Oral Cavity Squamous Cell Carcinoma
Development and Validation of Machine Learning Models for Predicting Occult Nodal Metastasis in Early-Stage Oral Cavity Squamous Cell Carcinoma
Importance
Given that early-stage oral cavity squamous cell carcinoma (OCSCC) has a high propensity for subclinical nodal metastasis, elective neck dissection has become standard practice for many patients with clinically negative nodes. Unfortunately, for most patients without regional metastasis, this risk-averse treatment paradigm results in unnecessary morbidity
Objectives
To develop and validate predictive models of occult nodal metastasis from clinicopathological variables that were available after surgical extirpation of the primary tumor and to compare predictive performance against depth of invasion (DOI), the currently accepted standard
Design, Setting and Participants
This diagnostic modeling study collected clinicopathological variables retrospectively from 7 tertiary care academic medical centers across the US. Participants included adult patients with early-stage OCSCC without nodal involvement who underwent primary surgical extirpation with or without upfront elective neck dissection. These patients were initially evaluated between January 1, 2000, and December 31, 2019.
Exposures
Largest tumor dimension, tumor thickness, DOI, margin status, lympho-vascular invasion, perineural invasion, muscle invasion, submucosal invasion, dysplasia, histological grade, anatomical subsite, age, sex, smoking history, race and ethnicity, and body mass index (calculated as weight in kilograms divided by height in meters squared)
Main Outcomes and Measures
Occult nodal metastasis identified either at the time of elective neck dissection or regional recurrence within 2 years of initial surgery
Results
Of the 634 included patients (mean [SD] age, 61.2 [13.6] years; 344 men [54.3%]), 114 (18.0%) had occult nodal metastasis. Patients with occult nodal metastasis had a higher frequency of lymph vascular invasion (26.3% vs 8.1%; P < .001), perineural invasion (40.4% vs 18.5%; P < .001), and margin involvement by invasive tumor (12.3% vs 6.3%; P = .046) compared with those without pathological lymph node metastasis. In addition, patients with vs those without occult nodal metastasis had a higher frequency of poorly differentiated primary tumor (20.2% vs 6.2%; P < .001) and greater DOI (7.0 vs 5.4 mm; P < .001). A predictive model that was built with XGBoost architecture outperformed the commonly used DOI threshold of 4 mm, achieving an area under the curve of 0.84 (95% CI, 0.80-0.88) vs 0.62 (95% CI, 0.57-0.67) with DOI. This model had a sensitivity of 91.7%, specificity of 72.6%, positive predictive value of 39.3%, and negative predictive value of 97.8%.
Conclusions
Results of this study showed that machine learning models that were developed from multi-institutional clinicopathological data have the potential to not only reduce the number of pathologically node-negative neck dissections but also accurately identify patients with early OCSCC who are at highest risk for nodal metastases.
Relevance to Healthcare Field
According to the American Cancer Society, Oral Cavity Squamous Cell Carcinoma (OCSCC) survival rate depends on the location and stage. While the localized Lip SCCC survival rate is 93%, and the Oropharyngeal SCC survival rate is 59%, the involvement of lymph nodes and metastasis will reduce the survival rate up to 10%. The Elective Neck Dissection (END) has been used for decades to improve survival rate, but it has been proven that it can bring negative outcomes when used on patients with a negative predictive value (NPV) of having a disease progression based on depth of invasion (DOI) and tumor thickness. The purpose of this study was to effectively calculate the risk of the patients developing vascular/node involvement and determine if the END is needed or not, thus avoiding unnecessary costs and side effects for patients who did not need it. The XGBoost machine learning model used more variables to determine treatment management: lymph vascular invasion (LVI), perineural invasion (PNI), grade, DOI, age, race, margins, and size. LVI and grade resulted as the criteria determining the biggest NPV, which is useful information for patients and specialists as they strive to determine cost-effective and time-effective treatment.