In this study we identified behavioral and neural characteristics of physicians who were adept at learning from experience. High performers learned from both successes and failures, and made smaller rule adjustments after feedback. Conversely, low performers learned disproportionately from successes, and made larger rule adjustments. Nearly half of the physicians performed at chance levels or worse, even after 64 training encounters. This result is particularly striking given that the difference in efficacy between the two treatments was intentionally made very large, at 75% versus 25% when diabetes status was taken into account. For comparison, the difference in real-world mortality between placebo and combined aspirin/streptokinase treatment for acute MI is only 13.2% versus 7.2% 
. The suggestion from these results is that for most physicians, clinical experience alone may be inadequate for forming reliable heuristics about most real-world treatments, even when the differences in efficacy are large.
Spurious learning was also widespread in this study. On behavioral measures, more than two-thirds of physicians robustly incorporated spurious associations into their treatment algorithms. Overall, subjects were nearly twice as likely to invent a spurious rule as to detect the correct one. The high rate of spurious rule formation, in conjunction with the tendency to ignore failures, suggests how different experts might come to disagree vehemently about which factors are most relevant to decision-making, based on their personal experiences. The implication is that experiential learning alone is likely to capture irrelevant as well as relevant factors for guiding decisions, even among experts working within their domain of expertise.
Notably, years-of-experience was as strong a predictor of poor performance as were the number of spurious rules (r
0.4 for each). Years-of-experience was also moderately (but not significantly) correlated with aggregate speed of learning (r
.24). The poorer performance of more experienced physicians on this task could conceivably reflect differences in training, temperament, or cognitive style. Although the present study was not designed to explore these possibilities in detail, understanding the basis of the experience-dependent decline in performance would be an important area for future investigation.
On neuroimaging, high- and low-performing subjects showed different patterns of activity in the inferior parietal cortex, and right DLPFC: brain areas with well-established roles in salience representation, associative learning, and the coding of prediction error 
. Low performers, who included irrelevant factors in their algorithms, showed stronger prefrontal and parietal activation for successes than for failures. Conversely, high performers showed stronger prefrontal and parietal activation after treatment failures than successes. The profile of activation in inferior parietal areas related to attention and salience 
similarly suggest that while low-performers pay special attention to successes, high performers attend more to failures during learning.
The right DLPFC has a well-established role in learning causal relationships 
. The present study extends these previous findings to medical learning and decision-making in expert physicians, participating in a medically-framed decision-making task. While the present study did not detect significant correlations between unsigned prediction error and DLPFC our findings are consistent with the underlying hypothesis that the right DLPFC drives rule readjustment 
. Specifically, the differential activation of the area after successes in low performers, and after failures in high performers, does reflect the learning biases that characterize the two groups (). The present study suggests that engaging these mechanisms following only successful predictions leads to inaccurate rule formation, and poorer predictions in the future, as spurious rules progressively accumulate.
Notably, the profile of DLPFC activation in low performers on this task bears a striking similarity to that seen during false-belief formation in pathological settings, such as in individuals with psychotic illness 
or ketamine-induced psychosis in healthy subjects 
. Thus, the existing literature on the role of the DLPFC in associative learning may have important implications not only for psychiatric patients, but also for the medical decision-makers who treat them.
If learning preferentially from successes is such an ineffective strategy, then why is it so pervasive even among experts operating in their domain of expertise? Here, the reward value of successes offers one possible explanation. Success-chasing, as measured by learning asymmetry, correlated significantly with anticipatory activity in the left NAC at trial onset (). Notably, both this anticipatory activation and activation subsequent to outcome revelation showed significant differences between successful and unsuccessful trials in complete success-chasers (those with learning asymmetries of 1). In contrast, those who learned more from failures (with learning asymmetries less than 0) showed no anticipatory activation at all () and no significant differences in activation between successes and failures either before or after outcome was revealed.
These results support an interpretation of confirmation-bias among success-chasers: the confirmation of an expected reward leads to increased signal in the NAC in these subjects. This profile of activation is particularly interesting in light of evidence showing that activation in the NAC correlates with susceptibility to the placebo-effect 
, another example of confirmation bias in a completely different context. Whereas in the placebo effect the expectation of and effective treatment can lead to the alleviation of symptoms, here the expectation of reward appears to be amplifying each decision-makers response to reward both physiologically in the NAC itself, and behaviorally in there adjustments of their underlying beliefs.
Why would success-chasing be so prevalent despite its drawbacks? The ventral striatum and nucleus accumbens is known to play a key role in motivation and the prediction of a wide variety of rewards: juice, consumer goods, monetary gains, and gains in social reputation 
. The anticipatory activation of this region among asymmetric learners in the present study suggests that success-chasers may be excessively motivated by the reward value of a successful outcome. In their efforts to maximize successful outcomes during training, success-chasers may be paradoxically sabotaging their ability to learn effectively from past experience.
Taken together, the behavioral and neuroimaging results suggest that success-chasing and confirmation bias may underlie the relative pervasiveness of premature, asymmetric learning and the resultant poor performance of the majority of physician subjects in the present study. The general human bias towards confirmation over disconfirmation in hypothesis-testing has been extensively documented in a variety of non-medical contexts, such as the Wason Card Task 
. Conversely, the necessity for disconfirmation learning in empirical investigations is a key principle identified by the philosopher of science, Karl Popper 
. Conceivably, providing medical professionals with formal training in disconfirmation learning could improve their ability to learn effectively from clinical experience in real-world settings. Exploring this possibility would be an important area for future research.
In conclusion, the results of this study show distinct patterns of learning, both behaviorally and neurally, between effective and ineffective learners among physicians making decisions in a medically framed learning task. The tendency to chase successes and ignore failures provides a simple computational model of how spurious beliefs might be formed, and how different individuals seeing similar data might learn very different sets of associations. The neural differences observed could conceivably be developed into useful biomarkers for essential differences in individual learning styles. These may in turn prove useful in identifying those individuals who can resist the impulse to chase successes, and hence learn most effectively from experience. Finally, we note that although this study focused upon the specific case of medical decision-making, the findings may be also be relevant to many other fields in which experts must make high-stakes decisions by drawing upon personal experience.