European Urology

The knowledge necessary to practice medicine has evolved over generations by the recognition of clinical features, the identification of responsible pathologic changes, and the observation of responses to treatments. Whilst this knowledge is based in science, the practice of medicine is an art. Each stage of a patient's journey requires the evaluation of many variables to produce a single outcome. The treatment of localised prostate cancer, for example, depends on the cancer, coincidental pathology (eg, lower urinary tract symptoms from benign prostatic hyperplasia), the patient, the patient's social environment (eg, his friend chose active monitoring), probable outcomes from each treatment, and many other variables (mostly lacking an evidence base). It takes years to learn the knowledge required to address this complexity and a lifetime to perfect decision making.

The human brain evaluates multiple factors with apparent ease but is prone to “system errors”. These errors include use of incomplete knowledge, recall bias, and decisions that are often inappropriately affected by recent events (eg, problems with last week's operation). It is likely that machines are better than humans at evaluating repetitive, rule-based decisions. Machines can be automatically updated to include the latest, most complete knowledge; do not suffer recall bias; compute the same outcome from the same inputs every time (not adjusted by a “hunch”); and are not affected by recent performance. But there is a problem: To create machines to solve questions, one must understand the rules that govern the answer to these questions and be able to measure the variables that determine the answers. For many questions in medicine, we do not understand the rules fully or cannot detect the most important variables. Additionally, life is full of chance events that are hard to incorporate into model-based analyses. To use the reliability of machine-based processing, we should apply machines to areas of medicine where the rules are better understood and the outcomes are less affected by chance. One such area is the prediction of cancer outcomes.

To the patient, the most important cancer outcome is cure. Cure rates incorporate three distinct results: (1) death from cancer, (2) cure of cancer, and (3) death from another cause whilst still affected by cancer. This first result is easy to determine from registry data. In 2009, around 40% of patients diagnosed with cancer died from that malignancy [1]. For bladder cancer, this rate is nearer 20%. The other two results are not easy to determine.

Registry figures merely look at diagnosis and death. For tumours affecting young patients, these may be sufficient to determine cure rates, but this is not the case for most cancers. The Bladder Cancer Research Consortium database shows that 44% of patients die from tumour following cystectomy and 16% die from other causes [2] and [3]. These proportions may change dramatically with tumour stage, histologic type, and disease natural history (eg, 63% of men with localised prostate cancer die from other causes with their disease in situ [4]). For most patients, however, survival is driven by the malignancy (around 60% either die from cancer or are cured). In this group, histologic criteria should be strongly predictive of clinical outcome and can be used as a rule base for a predictive model. Using pathologic criteria, models successfully predict outcome in 70–80% of patients [3] and [5]. They fail in 20–30% of cases because the training data were incomplete (ie, we do not know the most important parameters to measure) or included cases with outcomes that were not driven by cancer (ie, the patient died from other causes before an otherwise fatal malignancy progressed).

To improve modelling performance, one should reduce contamination (with patients whose outcome is not cancer driven) and include all variables that direct outcome. As clinicians, we know that many histologically identical tumours behave very differently. The progression rate of high-grade non–muscle-invasive bladder cancer varies between 25% and 75% [6], so there must be molecular factors that direct outcomes in tumours that look identical to the pathologist. To find these molecular factors, various experiments can be used, including gene expression microarrays [7].

Microarrays are molecular tools that can measure thousands of genes or molecules simultaneously. They can reveal new insights into biology and can categorise previously uncategorisable diseases. Microarrays, however, generate huge datasets that are difficult to analyse (eg, 32 000 genes per sample). We must learn to manage these datasets, as microarrays are modest in size when compared to current molecular technologies. Second-generation sequencing machines, for example, now generate 9 million results per sample.

The rapid progression of technology is best illustrated using DNA sequencing. Sequencing of the human genome was started in 1990 and took 13 yr to complete at a cost of $3 billion (using first-generation sequencing technology). In 2008, James Watson's genome was sequenced in 4 mo and cost just $1.5 million [8]. In 2010, the next (third) generation of sequencing technology will be introduced. It will produce more data for less cost and can even sequence single cells. Within a decade, it is possible that we will be able to use the entire genome of an individual patient's cancer to predict prognosis and response to treatment, but only if we can understand the information obtained.

With these technological advances in mind, we have applied artificial intelligence to microarray data [9]. As discussed by DeGraff [10], we validated our outcomes in a new tumour cohort using immunohistochemistry (a simple molecular technique, available in every hospital). We found that our predictive panel stratified outcomes in a second independent patient cohort for 66% of patients (concordance). This percentage is lower than we hoped but bears remarkable resemblance to the proportion of bladder cancer patients whose outcome is driven by malignancy. We thank Dr. DeGraff for his kind words relating to our work and encourage others to explore the field of artificial intelligence.

Whilst statistical methods have withstood the rigors of time, are understandable, and are founded in thorough mathematical proofs, the rapid progress in molecular biology and the needs of the clinician require expansive tools not confined to linear relationships, if we are truly to improve patient care.

Conflicts of interest

The authors have nothing to disclose.