Biases in Human Prediction

Medical Prediction

One major type of medical prediction is the diagnosis of patients' disease. Diagnosis involves gathering and integrating evidence, testing hypotheses, and assessing probabilities. This requires that a clinician be able to generate accurate predictions from incomplete data about the underlying cause(s) of the patient's symptoms. For example, the symptom “pelvic pain” might be caused by a urinary tract infection, a sexually transmitted infection, or by cancer, among other possible diagnoses. A physician who sees a patient with this symptom must accurately predict the likelihood of multiple possible underlying causes to effectively gather evidence (i.e., ask about other possible symptoms and order appropriate tests), cognitively integrate that evidence, and determine the most probable diagnosis.

Once the physician has made a diagnosis, he or she must, along with the patient, make another medical prediction when they decide together on a treatment decision. Selecting the optimal treatment from multiple options requires that a clinician be able to predict which treatment will provide the patient with the best possible health outcome, accounting for both positive and negative effects. For example, a patient with localized prostate cancer has multiple treatment options available, including surgery, radiation therapy (of two types), hormone deprivation therapy, and surveillance. To make a treatment recommendation, a physician must predict the patient's response to various treatments, both in terms of disease control and potential burden from treatment side effects. The physician must also consider the patient's overall health, comorbidities, resources, social support, and preferences for possible health states.

Physicians also make medical predictions when necessary to provide prognoses, which are predictions of the likely duration, course, and outcome of a disease based on the treatment chosen. This is particularly important in diseases, such as terminal cancer, where patients and their families wish to form appropriate timelines for goals of care and to have access to certain types of care, such as hospice, when they would most benefit from them. Unfortunately, as Nicholas Christakis has shown, prognosis is particularly difficult in emotionally intense situations such as this.

Given the centrality of accurate predictions to medical decision making and the common assumption that medical training improves physician's decisions, it is disheartening that research has repeatedly shown that physicians' medical predictions are as susceptible to cognitive biases as others are in nonmedical domains. The mistakes are systematic, not random, errors that are likely due to the difficulty of the prediction task combined with human psychology. Thus, these biases are not significantly reduced by current medical training. As Reid Hastie and Robyn Dawes argue, one of the most persistent of these biases is overconfidence concerning one's predictions. The danger of overconfidence is that one cannot begin to correct other biases affecting the quality of one's predictions; simply recognizing their existence is something that overconfidence prevents. For example, an overconfident surgeon might regularly predict better surgical outcomes for his or her patients and perform surgeries on patients who are poor candidates for surgery. This overconfidence bias will go uncorrected because it is unrecognized as a systematic error.

...