When a patient is diagnosed with cancer, doctors and families want to know what to expect, because the prognosis shapes treatment choices. This study used machine learning to predict survival for thyroid cancer patients.
The problem
Predicting prognosis is essential for choosing a course of treatment. Many cancer survival models exist, but results can vary widely depending on how much data is used and which factors are chosen.
What the research did
We used data from SEER (Surveillance, Epidemiology, and End Results), a major US cancer registry. We selected features carefully, measured correlations and identified 17 crucial factors. We then trained a group of classification algorithms and showed how results change with different data sizes and factors.
What we found
The models predicted survival with accuracy ranging from 97% to 99%, and were also evaluated on F1-score, precision, recall and AUC. To our knowledge, the model was more accurate than those in earlier studies.
Who benefits
- Oncologists, as a decision-support aid when discussing prognosis
- Patients and families, through clearer information
- Health researchers studying cancer outcomes
Limits and what comes next
The models are based on US registry data and may not transfer to other countries without retraining. Survival predictions are probabilities, not certainties, and should only be used alongside a doctor’s judgement.
The team
This research was led by Saadat Mehmood Alhashmi (University of Sharjah, UAE) with colleagues at Daffodil International University including me, Mohamed Hashem (University of Sharjah) and Nirase Fathima Abubacker (Asia Pacific University, Malaysia).
Frequently asked questions
Can AI predict cancer survival?
Machine learning can estimate survival probabilities from patient data, as an aid to doctors.
What data was used?
The US SEER cancer registry.
Should patients rely on these predictions?
No. They are a support tool; prognosis should always be discussed with a doctor.