Machine-Learning Models Can Help Detect Early-Stage Cancer
A new study suggests that machine-learning models can predict occult nodal metastasis in patients with a type of early-stage oral cavity cancer with more accuracy than standard methods.
Source: Getty Images
By - A recently published diagnostic modeling study published in JAMA Network Open successfully developed machine-learning algorithms to predict occult nodal metastasis in patients with early-stage oral cavity squamous cell carcinoma (OCSCC), a type of oral cavity cancer.
OCSCC has significant rates of subclinical nodal metastasis, which has led to elective neck dissection (END) becoming the standard practice for treatment. However, for patients without regional metastasis, this can result in unnecessary morbidities. Regional metastasis associated with OCSCC may also increase cancer recurrence rates and decrease survival rates.
For the purpose of this study, researchers developed both a predictive model and classification algorithms to be used for analysis together so that multiple clinical and pathological variables could be considered, rather than relying on only one of these groups of variables.
Researchers developed the predictive model based on tumor depth of invasion (DOI), the current standard for predicting occult nodal disease, because of the demonstrated association between the two. Using this model, people with any tumor with a depth at or greater than 4 millimeters are predicted to have occult nodal disease. In contrast, any tumor below that threshold indicates there is no disease.
In addition to the predictive model, researchers also developed classification algorithms to predict pathological node (pN) metastasis. Clinical variables, such as age, sex, race, and ethnicity, were utilized alongside pathological variables, such as DOI and histological grade.
By using both the predictive model and the classification algorithms, the machine-learning algorithms developed could correctly predict nodal metastasis and potentially reduce ENDs for patients with pathologically negative lymph nodes. The algorithms predicted nodal metastasis with significantly greater accuracy than models based on DOI alone. The study indicates that the classification algorithms misclassified only 8.3 percent of patients with positive pN while using DOI alone resulted in a 37.5 percent misclassification rate. Similarly, the predictive model misclassified 27.4 of patients with negative pN compared with a 38.7 percent misclassification rate while using DOI alone.
The success of this predictive model indicates the potential for machine learning to improve risk estimation for early-stage OCSCC patients and reduce ENDs for pathologically node-negative patients. Improved risk estimation may enhance outcomes for those at the highest risk for nodal metastases, while reductions in ENDs for pN-negative patients have the potential to eliminate unnecessary morbidity and decrease costs overall, according to the study.
Machine learning is seeing some potential for use in cancer care. A recent study shows that machine-learning algorithms can help clinicians detect patients at high risk for colorectal cancer, enabling the care team to intervene and schedule colonoscopies for these patients.
But, while machine learning in cancer care is a promising innovation in accelerating care and treatment, experts suggest that artificial intelligence (AI) in healthcare requires an evidence-based development and deployment approach.
