Prevention is better than cure. This is something that everyone must have heard from doctors and even your parents. Today, researchers are trying to utilize machine learning in cancer risk prediction. This is important because it can offer some sort of alarm before a tragedy occurs. It allows doctors to provide preventive measures against the disease. However, is it currently on par with the standards used in medicine? We can answer this by perusing the thousands of research articles about cancer and artificial intelligence. This discussion will help in summarizing the current state of artificial intelligence used for cancer risk prediction.
Cancer and Risk Prediction
Risk prediction for cancer is not new. There have been several attempts for this that do not include artificial intelligence (AI). Previously, the most used implementations were statistical analysis tools such as logistic regression. However, AI proves to be more efficient when it comes to more significant amounts of data.
Feeding data into an AI preemptively will make the final model faster. Through this, the model learns how to predict the outcome of the data fed into it. Nevertheless, these models must still follow certain specificity and sensitivity standards.
These terms will be important in our discussion as they describe a model’s effectiveness. Sensitivity tells us how well a model can predict a positive result as opposed to false negatives. On the other hand, specificity tells us how well a model can predict a negative result as opposed to false positives. To combine these two descriptors, scientists used the Area Under the Receiver Operating Characteristic Curve (AUC) score. A good AUC score is between 0.8 to 1.0.
As a specific case of AI in cancer risk prediction, we can look to this 2021 comparative study. In this study, researchers compared logistic regression analysis against AI. They tested these models to predict cancer risk through laboratory blood tests. Results showed that logistic regression and AI were comparable in a cohort of 6,592 profiles. However, AI had an edge on validation tests of 1,368 profiles in terms of AUC and sensitivity/specificity.
Researchers conducted this comparative study with no specific cancer in mind. They tested blood profiles and predicted the risk of several types including prostate, lung, ovarian, and stomach cancer. Of course, results will pale in comparison to a specialized model for each type of cancer. Because of this, we will proceed to each of the deadliest cancers in need of an early prognosis.
Lung Cancer Risk Prediction
As previously discussed, there are earlier models and recommended guidelines for cancer risk predictions. Authorities such as government-backed medical institutes often state these guidelines. A 2020 study compared these guidelines to an AI algorithm called the lung cancer prediction convolutional neural network.
In this study, researchers collected samples of pulmonary nodules to predict lung cancer. They collected 1,397 nodules in 1,187 patients and used a national guideline and AI to predict cancer risk. The comparison with the AI algorithm was against UK guidelines from the Brock University model.
Results showed that the AUC score for the AI algorithm was 89.6% while the Brock model had 86.8%. However, the AI algorithm had a better sensitivity rating. Additionally, the AI algorithm found more nodules below the lowest cancer nodule score of 24.5%.
Colorectal Cancer Risk Prediction
As early as the T1 colorectal cancer stage, the tumor may require resection. Doctors may recommend surgery even with a low incidence of metastasis. If we were to predict the accurate chance of metastasis at the earliest, we could avoid this early need for surgery.
This is the intention behind the 2020 gastroenterological research that uses AI for cancer metastasis prediction. The study collected data from a cohort of 3,132 colorectal cancer patients from six hospitals. Results showed that the AI model had an AUC score of 0.83. In comparison, the US guidelines only had an AUC score of 0.73.
Pancreatic Cancer Risk Prediction
This type of cancer typically proliferates undetected due to poor prognosis. At the same time, it has a very low 5-year survival rate. An accurate risk prediction can encourage preventive measures against the aforementioned outcomes.
To battle this dilemma, scientists are utilizing AI against a very large number of data. In this 2021 multi-national study, researchers used AI on the Danish National Patient Registry and the Mass General Brigham Healthcare. The dataset had a total of over seven million patient records to train the AI. There were multiple tests, and the best model had an AUC score of 0.88 for cancer occurring within 36 months after risk assessment.
Breast Cancer Risk Prediction
This disease causes the most deaths of any cancer among women. Researchers hope to reduce this by introducing AI into cancer prediction. Specifically, they can also incorporate AI into existing models to improve its effectiveness.
Scientists performed this incorporation in this 2019 research. They integrated logistic regression, linear discriminant analysis, and neural network models into the Breast Cancer Risk Prediction Tool (BCRAT). These integrations saw an improvement in the five-year breast cancer risk prediction compared to BCRAT alone.
Aside from these existing models, standalone AI algorithms for breast cancer risk prediction do exist. The following include the Support Vector Machine, Decision Tree, Naive Bayes, and k Nearest Neighbors.
Prostate Cancer Risk Prediction
While prostate cancer is still one of the deadliest cancers, there hasn’t been much progress in incorporating AI in its predictions. This may be because prostate cancer has a high survival rate due to hormonal therapy advancements. In addition, it also has a relatively easier diagnosis compared to those discussed above.
As for current developments in prostate cancer risk prediction, this 2018 study highlights the progress. Researchers in the study used data from the National Health Interview Survey conducted. The AI model only had an AUC score of 0.73 for the training set and 0.72 for the validation set. Although, there may be applications for this as it still has a high specificity in its predictions.
