{"doi":"10.1111/bju.15951","title":"Do not treat Bill Gates for prostate cancer! Algorithmic bias and causality in medical prediction","abstract":"Chase et al. [1] recently published in the BJUI a life-expectancy prediction model for patients with prostate cancer. They expressly recommended this be used to guide treatment decision-making for patients, and even provide an easy-to-use app for doing so. It is the authors’ explicit intention that urologists use their prediction model to decide who should and should not receive curative treatment depending on whether life expectancy is >10 years. The underlying principle is completely sound. Life expectancy is indeed a critical determinant of prostate cancer treatment and many current approaches, such as using social security tables, or informal clinical assessment, are known to be invalid [2]. The problem with the Chase et al. [1] model is that it includes education and marital status. If you take two men with identical tumours, and who are also the same age and have the same comorbidities, you might recommend one but not the other to undergo curative treatment based on his higher level of education, or because he is married. Indeed, one might imagine a situation where a man is scheduled for a radical prostatectomy, but then his wife leaves him and the urologist, after using the app to make a quick calculation, tells the patient that he is no longer eligible for surgery. Perhaps even more problematic, use of the prediction model would systematically discriminate against Black Americans. Centuries of racism has left Blacks less likely to be well-educated or married. So randomly select a Black man and a White man of the same age, cancer and comorbidity profile, there is a good chance that the White man will be referred for surgery or radiotherapy whereas the Black man is denied curative treatment. This can only worsen the severe and chronic racial disparities in prostate cancer outcome that continue to blight American healthcare. To their credit, the authors are not unaware of the problem of algorithmic bias, even citing the seminal work of Kent and Paulus [3]. My own view, however, is that their arguments are weak. Take, for instance, their assertion that ‘<5% of patients … would have had their prediction appreciably changed by a modification to their marital status or educational attainment’. This is far from reassuring. Let us be 100% clear: no-one should be denied treatment because they are unmarried or never got a degree. As it turns out, Bill Gates, an unmarried college drop-out billionaire, might be one of the men who would be given a different prediction because the Chase et al. [1] model includes education and marital status. Indeed, the Bill Gates problem tells us exactly why we should avoid correlates of social determinants of health in prediction models. Causality is not something we normally worry about when making predictions: if it predicts, it predicts. For instance, grip strength in older patients predicts life expectancy and so might be used as a simple, in-office test to help guide shared decision-making [4]. No one thinks that poor grip causes an early death, but that is irrelevant. However, it is rather different for predictors that have social consequences. We must think about causality if our model might exacerbate disparities. Educational and marital status are predictive of life expectancy in patients with prostate cancer because they are associated with social support and access to care and so, as in the case of Bill Gates, we are better off examining a patient's level of social support and access to care than to simply ask about their education and marital status. The numerous medical prediction models that directly include race as a predictor [5] should similarly explore causality. In many prostate-cancer models, Black race is a correlate for genetic differences that are yet to be fully elucidated, and it seems appropriate to include race until better markers of genetic risk are available. In other models, it seems that race is used simply because it is predictive and that will often be because of r","journal":"British Journal of Urology","year":2023,"id":387389,"datarank":0.0,"base_score":0.0,"endowment":0.0,"self_citation_contribution":0.0,"citation_network_contribution":0.0,"self_endowment_contribution":0.0,"citer_contribution":0.0,"corpus_percentile":null,"corpus_rank":null,"citation_count":2,"citer_count":0,"citers_with_citation_signal":0,"citers_with_endowment":0,"datacite_reuse_total":0,"is_dataset":false,"is_dataset_confidence":0.9487,"is_data_producer":false,"deposit_databanks":null,"is_oa":true,"file_count":0,"downloads":0,"has_version_chain":false,"published_date":"2023-01-01","fair_score":null,"fair_percentile":null,"algorithm_id":"datarank_citation_only_1hop_v6","ranking_scope":"data_only","authors":[{"id":72212,"name":"Andrew J. Vickers","orcid":"0000-0003-1525-6503","position":0,"is_corresponding":true}],"reference_count":6,"raw_metadata":null,"created_at":"2026-07-19T01:18:04.706547Z","pmid":"36716733","pmcid":null,"fwci":null,"citation_percentile":null,"influential_citations":0,"oa_status":null,"license":null,"views":0,"total_file_size_bytes":0,"version_count":0,"fair_f":null,"fair_a":null,"fair_i":null,"fair_r":null,"fair_zscore":null,"fair_rationale":null,"fair_model":null,"fair_agent_version":null,"fair_fulltext_source":null,"fair_has_llm":null,"fair_computed_at":null,"clinical_trials":[],"software_tools":[],"db_accessions":[],"linked_datasets":[],"topics":[]}