What’s new in cancer: Prognosis and prediction

A cancer diagnosis is a fact wrapped in doubt. The famous physician William Osler said that medicine is a science of uncertainty and an art of probability. Even before a diagnosis is made we know that people have doubts and worries about timelines, delays in the system, and the results of tests. When a diagnosis eventually arrives it triggers more uncertainties about outcomes. Clinical protocols, research data and past experience can help explain to a person affected by cancer the probabilities of what should happen next, but guarantees are rare. Statistics from populations may not help an individual; for example, I know the average age at death for men in the UK is 82.3, but that doesn’t tell me how long I will live, an event that will be influenced by multiple individualised factors, some known yet many others unknown.

Sharing accurate data on the likely treatment impact is an important part of shared decision making. Patient Decision Aids (PDAs) are tools that help this process. In prostate and breast cancer there are excellent PDAs in the form of the Predict tools for patients to use with their healthcare team. These online calculators use large datasets from UK patients to give an estimate of outcomes. Predict Prostate allows comparisons of the different forms of treatment for early prostate cancer (radiotherapy, surgery and watchful waiting) for both treatment efficacy but also likely side effects for bladder, bowel and sexual (erectile) function.  The tools can be refined as more data becomes available and a new version of the Predict Breast tool is currently being piloted to include radiotherapy benefits, and harms of smoking and impact on the heart from treatment.

These models use NHS data matched to patient’s tumour characteristics based on standard models we have used for decades (size, grade, spread etc). In an era of genomic medicine, it is hoped that more accurate prognosis and prediction might be achieved using more detailed analysis of the tumour. In May, NICE approved the use of Genomic Expression Profiling tests for people with early breast cancer that had spread to lymph glands. Traditionally lymph gland spread has been a strong indicator to recommend chemotherapy but these genomic tests help predict those people who will not benefit from chemotherapy, sparing a great deal of distress and toxicity.

Looking deeper at genomics has led to systems which help classification and individualisation of prognosis in multiple different cancers including myeloma, prostate and pancreatic cancer. Macmillan are currently surveying healthcare professionals to understand barriers and opportunities in biomarker testing. As genomic medicine accelerates our knowledge, there is a race to find biomarkers that will guide clinicians and inform patients but in all the noise it can be hard to find the signal. A review of breast biomarkers published in July identified 4,500 articles describing 2,347 biomarkers of which only 23  (0.94%) are currently recommended for clinical use. This is a problem of redundancy identified across the field of prognostic biomarkers. It’s hoped that the Deep Learning tools of Artificial Intelligence can comb through vast datasets to find useful prognostic signals. There are multiple efforts happening globally to take genomic data and, where possible, combine molecular data from precision medicine with real world outcomes from patient databases. When cancer treatments are increasingly expensive and have difficult toxicities, it becomes more important than ever to find clues as to which people will benefit the most.

Identifying benefit is important but the primary medical ethic is ‘primum non nocere’ (first do no harm) and so predicting which therapies will not help is an imperative. This is arguably more important in palliative treatments where life may be limited, and any time spent having ineffective or harmful treatment is precious time wasted. There have been multiple efforts to design predictive scores for palliative radiotherapy with recent efforts confirming that there is still no clear and reliable scoring system. Machine learning has also been used to predict benefit for chemotherapy in stomach cancer. This is important work that continues, but the lack of a consistently reliable system shows how difficult the work is. Future scoring systems need to incorporate data sets that are representative of our populations and all patient groups,

Arguably the most difficult prognostication (both emotionally and practically) is predicting mortality in people approaching end of life. Multiple tools and algorithms have been developed to try and answer this eternal question and none work with the certainty we may want. We will always have some degree of uncertainty. Helping people affected by cancer manage uncertainty (as with Macmillan resources) may ultimately be a more fruitful endeavour than chasing an elusive and unreliable answer. A good conversation that reflects individual values is more helpful than a prognostic algorithm.  Science will continue to narrow the uncertainty gap, but we will always need to support people as they navigate the unknown.