Guest Editorial: The Health Benefits of Bayesian Reasoning

One of the most prevalent and harmful epidemics both in contemporary medicine [25, 26] and orthopaedic surgery [7, 11] is not caused by an infectious disease, but rather by physicians like us. We speak, of course, of overdiagnosis and overtreatment. Fortunately, incorporating practical concepts from Bayesian reasoning into our decision-making framework may help us overcome the common errors that contribute to these problems [2, 10, 20, 27]. The Bayesian framework is not a parlor trick for statisticians; it is simply the formalized version of the same iterative, context-sensitive reasoning surgeons have always employed. Bayesian reasoning is the process of updating our knowledge by interpreting new evidence within the context of what we already know [2]. Orthopaedic surgeons use it when ruling out serious conditions that have uncommon symptoms but a high risk of danger—like compartment syndrome, necrotizing fasciitis, or compressive median neuropathy—and when deciding whether imaging would be helpful or meddlesome for conditions that have common symptoms but a low probability of danger, like common back or knee pain [27]. In other words, effective Bayesian reasoning helps clinicians manage uncertainty [12]. Bayesian reasoning does this by specifically addressing probabilities. Thinking in terms of probabilities rather than certainties helps us remain thoughtful in our actions and protects against both overreaction and complacency. For instance, a clinician using Bayesian reasoning doesn’t consider whether the diagnosis of rotator cuff tendinopathy is accurate or not; instead, a clinician considers how likely the patient’s symptoms are related to degeneration in the rotator cuff tendons. In practice, probabilities rarely reach 0% (we’re positive the diagnosis is absent) or 100% (we’re positive it’s present). Therefore, high-probability diagnoses—where a specific pathophysiology is the most likely cause of symptoms—allow us to act confidently while remaining open to new information. If new evidence changes the diagnosis, we do not consider the initial assessment a misdiagnosis; rather, we remain confident that, over time, new information will help us identify any important pathophysiology. Bayesian reasoning also keeps us mindful of the potential for overdiagnosis, overtreatment, and iatrogenic financial or psychological harms, because some “updates” to probabilities are relatively inconsequential. For instance, a person aged 40 or older with anterolateral shoulder pain during elevation or internal rotation and full passive shoulder motion is likely to have rotator cuff tendinopathy (with or without biceps tendinopathy). Confirmatory imaging can only modestly increase this already-high pretest probability, and the limited potential for benefit is outweighed by the potential for harm, as diagnostic imaging may draw unhelpful attention to incidental findings, such as acromioclavicular osteoarthritis [17] or signal changes in the superior labrum [9]. Both of these conditions are nearly universal with age and rarely cause symptoms, yet too often are treated with unnecessary surgery [19, 24]—which is, by definition, harmful [15, 16]. On the other hand, some updates to probabilities are informative [18]. For instance, if a person over age 40 with anterolateral shoulder pain has limited passive motion, the probability of rotator cuff tendinopathy is reduced whereas the odds of glenohumeral arthritis or adhesive capsulitis become reasonably high. If radiographs confirm glenohumeral arthritis, the probability of rotator cuff tendinopathy being the sole source of symptoms approaches zero. In this case, the information gained from the physical examination and imaging are of greater consequence and utility (Fig. 1).Fig. 1: A Fagan Nomogram is a quick, visual method for estimating posttest probability based on pretest probability (knowledge, symptoms, signs) and the likelihood ratios (LR) of the specific test in similar settings. Blue–Rotator cuff tendinopathy. A 55-year-old man with anterolateral shoulder pain with elevation and internal rotation and normal shoulder motion. The pretest probability of rotator cuff tendinopathy prior to imaging is high (clinician estimate 95%) based on the high prevalence of rotator cuff tendinopathy and the characteristic symptoms and signs. An MRI in this setting is likely to demonstrate signal changes in the rotator cuff tendons and confirm the diagnosis (LR 10). Using the nomogram, the posttest odds of rotator cuff tendinopathy as the source of the symptoms increases to nearly 100%. But the diagnosis was already likely to be rotator cuff tendinopathy, and we don't need the certainty to be higher in order to guide the patient. The small difference between 95% and near certainty does not provide sufficient benefit to outweigh the costs, inconveniences, and potential harms (iatrogenic, financial, psychological) associated with the test. Green–Glenohumeral arthritis. If the person in the above scenario has limited passive motion, that creates a moderate pretest probability of glenohumeral arthritis or adhesive capsulitis, perhaps around 40% each. Radiographs can detect glenohumeral arthritis (LR 10). In this case, a radiograph could potentially increase the posttest odds of being able to specifically diagnose glenohumeral arthritis (with abnormal radiographs) or adhesive capsulitis (with normal radiographs) as the source of the symptoms to above 90%. This is an effective use of a diagnostic test because rotator cuff tendinopathy, adhesive capsulitis, and glenohumeral arthritis have different prognoses and treatment options. Red–Nonspecific wrist pain. There is a low pretest probability of any specific pathophysiology in a 30-year-old healthy woman with diffuse wrist pain, full passive motion, and normal radiographs. For this exercise, we need to consider a specific diagnosis. Let's set the pretest odds of scapholunate interosseous ligament (SLIL) rupture at 1%. Again, for the sake of the exercise, we'll put aside that MRI signal changes in the scapholunate ligament may be common incidental findings and can be difficult to associate with symptoms. Let's set the likelihood ratio relatively high (LR 10), which means that if there is scapholunate pathophysiology, the MRI will identify it. Using the nomogram, we see that in a setting of low pretest odds for a specific pathophysiology (1% odds of SLIL rupture), the posttest odds remain low (perhaps 10%) that the pathophysiology of the scapholunate ligament is the source of symptoms, even if the MRI shows scapholunate ligament signal changes. Treating the scapholunate ligament as the source of symptoms therefore has a high probability (about 90%) of representing overdiagnosis and overtreatment. This is probably the least intuitive and most impactful aspect of Bayesian reasoning—that acceptance of uncertainty may be healthier than “taking a look.” A color image accompanies the online version of this article.Perhaps the most useful rule of thumb is that scenarios with uncertainty—characterized by low pretest odds of any specific diagnosis—can be treacherous to our patients [27]. It’s not uncommon for a person to present with disproportionately intense, diffuse, or persistent symptoms, which can be associated with feelings of anxiety and fear of painful movement [3, 13, 23]. Protectiveness can also hinder the physical examination [28]. In a person aged 40 or older, the initial odds that the pain is from rotator cuff tendinopathy are high because the condition is nearly universal with age [22]. It is, by far, the most common cause of shoulder pain and the most common reason for a person to seek care for an upper extremity problem in general [5]. But what about a 30-year-old with diffuse pain and normal passive motion in the wrist? Nonspecific wrist pain is common [14]. The probability of any specific pathophysiology is low, and the potential for harm from overinterpretation and overtreatment of incidental MRI findings and harmless anatomical variations is much higher—half of all asymptomatic wrists have a ganglion cyst [6], articular disc abnormalities are common [4], as are incidental tendinopathies of the extensor carpi ulnaris [8] and flexor carpi radialis tendons [21]. There is a low probability that diffuse and disproportionate pain symptoms relate to any of these imaging findings, but they’re often treated with surgical interventions that might be considered expensive, dangerous placebos. In the absence of potentially dangerous pathophysiology such as infection, a Bayesian approach is to be comfortable with the uncertainty and absence of a definitive solution, accompany the patient over time, collaborate on alleviating concerns, and guide them through reorientating common misconceptions. Over time, this process may facilitate a clearer description of the symptoms and enable a more useful examination, which could increase the odds of a specific diagnosis. If so, you have created a circumstance in which a specific diagnostic test may have a meaningfully greater potential for benefit than for harm. When we move from bedside diagnosis to the interpretation of scientific reports, the same Bayesian logic applies. Asking, “How surprising is the result of this experiment I’ve just read, given what we already know?” is the essence of sensible scientific judgment. When specific effects are a priori plausible but uncertain, positive findings in a well-designed, placebo-controlled trial can meaningfully increase the probability that any observed alleviation of symptoms after treatment correspond with alterations in the pathophysiology. In contrast, a finding of little or no difference from placebo can meaningfully increase the probability that alleviation of symptoms is due to nonspecific effects, such as the placebo effect, regression to the mean, and the self-limiting symptoms of many diseases [10, 20]. When two active treatments are compared, at least one should have a high probability of specific benefit [1]. Otherwise, the results of the trial cannot increase or decrease the probability of specific effects. In the setting of an experiment that tests for unlikely results—for example, durable alleviation of symptoms or any degree of disease modification for a permanent degenerative condition such as osteoarthritis—the prestudy odds of specific effects are low. This setting requires stronger evidence to produce a convincing alteration in probabilities, as inadequate consideration of prior plausibility and study quality can lead to overinterpretation (Table 1). Bayesian reasoning also reminds us that a meaningful increase in the probability of relatively implausible results can only occur with reliable, trustworthy replication of the findings. Table 1. - Bayesian reasoning and the assessment of experimental evidence comparing two treatments Prestudy probability of meaningful difference between treatments An experiment comparing the long-term outcomes of… Potential increase in poststudy probability of meaningful benefit High Operative vs. nonoperative treatment of an unstable nonunion of a diaphyseal humerus fracture Minimal Intermediate Operative vs. nonoperative treatment of a > 10-mm supraspinatus tendon defect with good muscle quality Moderate Low Debridement of osteophytes vs. sham debridement of osteophytes Minimal Practically, this means we should treat research findings the same way we treat diagnostic tests—ask what the result adds to what we already know. Consider the study claim, the prestudy likelihood of that claim being true, and the strength of the study signal (effect size, precision, and risk of bias) [12]. Together, these determine whether the posterior probability (that is, the probability in one’s mind after reading the study that the contention is true) is high enough to change practice [27]. When the evidence for a plausible, novel intervention is strong and consistent, it is appropriate to consider it an option. In contrast, less-plausible effects should be considered unlikely in the absence of consistent findings in high-quality trials. In other words, a positive test with low pretest odds only modestly increases posttest probability; a positive study in a low-plausibility context rarely justifies a change in care. To make this concrete without resorting to cumbersome math, consider these rules of thumb (Table 2): Assess plausibility: Estimate initial odds of specific effects based on the mechanism of action, the pathophysiology, and existing evidence. Implausible effects (low prestudy odds) can only yield modest increases in poststudy probability, while genuine uncertainty about the efficacy of an intervention (moderate prestudy odds or clinical equipoise) is a situation in which new information (like a well-performed study) can increase knowledge (greater poststudy odds). Consider effect size and precision: Small differences and overlapping confidence intervals should yield only modest increases in posttest probability. Larger differences merit greater increases in poststudy probability, more so if those differences are reproducible (beware the outlier notable difference). Consider study quality when assigning probabilities: Design, blinding, comparison appropriateness, data quality, and conflicts of interest should affect how much a result should shift one’s beliefs. Replication and consistency: Consistent findings across multiple trials increase confidence; inconsistent results decrease it. Table 2. - A guide to Bayesian reasoning regarding experimental evidence Rule/consideration Potential to increase knowledge (poststudy probability) Limited Meaningful Assess plausibility: Estimate pretest odds of specific effects based on the mechanism of action, the pathophysiology, and existing evidence. Implausible effects (low prestudy odds) Genuine uncertainty (moderate prestudy odds, clinical equipoise) Consider effect size and precision Small differencesOverlapping confidence intervals Large differencesCaveat: beware of outliersa Consider study quality when assigning probabilities: Design, blinding, comparison appropriateness, data quality, and conflicts of interest should affect how much a result should shift one’s beliefs. Low quality experiment High quality experiment Replication and consistency Few experimentsInconsistent findings Multiple experimentsConsistent findings aOutliers may be associated with conflict of interest/ego/promotion, publication bias, or poor study quality. Making this perspective of Bayesian reasoning explicit may help surgeons and patients interpret evidence responsibly, reduce harm from overinterpretation, and better integrate new evidence into care.