Scratch N Sniff
Medical Case Study Analysis

Probabilistic Classification Learning With Corrective Feedback — a Peer Review

A peer review of Probabilistic Classification Learning With Corrective Feedback is Associated with in vivo Striatal Dopamine Release in the Ventral Striatum While Learning Without Feedback is Not Human Brain Mapping. This executive summary breaks down the impressive work for busy neuroscience professionals.

Explicit/conscious (Declarative) and implicit/unconscious (Procedural) forms of learning were considered part of two memory systems; the medial temporal lobe and the basal ganglia which are now known to interact cooperatively. A task that has been used to test this is a weather prediction task that involves incremental learning over trials where participants make predictions involving the particular arrangement of 1, 2, 3, or 4 possible tarot-like cards that display different patterns. Each card represents a different outcome (rain or shine) and the prediction is based on probability of .2, .4, .6, or .8 probability. Parkinson’s disease patients and those with amnesic patients were shown to have impairment relative to the controls over standard version patients. This was later refined to show that patients however were unimpaired when there was paired association therapy or where learning was observed and without corrective feedback vs when learning occurred with corrective feedback. The reasoning behind it was the learning was because patients recruited more basal ganglia to enhance the learning process. The impairments also were relative to the severity of the disease with less severe patients being unimpaired at weather prediction therapy.

More recent studies show that selective recruitment of the stratum during therapy provides more convincing evidence however no studies investigated in live patients the role of dopamine in modulation of paired association (PA) or corrective feedback during learning (FB). This study, using C-raclopride (RAC) positron emission tomography (PET) as a measurement tool, set out to show that corrective feedback (FB) and not paired association would bring about a significant decrease in striatal C-raclopride (RAC) binding relative to a control task. RAC binding is inversely related to the concentrations of endogenous synaptic dopamine at the time of scanning. To show comparisons, two scans were performed either in an active task or in a control task and all participants went through a RAC PET scan twice in a four-week period.


15 participants without any neurological order were recruited. All were screened for impairment or depression and then randomly assigned to two groups FB (corrective feedback) or PA (paired association). There was no difference between the two groups in terms of age or years of education, IQ, or MMSE scores. The material used was a set of four tarot cards projected on a computer screen with 400 training trials (more trials than usual to adapt to the scanning procedure). Using the 1, 2, 3, or 4 cards, there were 14 possible arrangements. Each card was associated with a specific weather pattern given the number of diamonds, squares, circles, or triangles on the card. Two cards were predictive of weather being good (one weakly and the other strongly) and two cards were predictive of the weather being bad (one weakly and the other strongly).

Each test subject was the same frequency of card arrangement but the order of those arrangements was randomized. In the corrective feedback group, the cards were viewed on screen for 7 seconds, and participants were asked to make a prediction on each trial using two-button responses to indicate a “thumbs up” or “thumbs down”. If a patient failed to make a response/prediction, no feedback was provided. The tasks were started 5 minutes before a tracer was injected and ended 5 minutes before a RAC PET scan was used to calculate and assess performance.

The PA or paired association group were shown the outcome of the arrangement with no classification required. Participants were asked if they had seen the card arrangements/outcome but there was not a timing constraint of the presentation. There was a two-second blank screen between combinations. The training also started 5 minutes before the injection of the tracer and ended 5 minutes before a RAC PET scan was used. A 10 min transmission scan was also performed prior to injections to correct for tissue attenuation of the gamma radiation. Scans were generated and simplified to estimate the tracer uptake and a standard region of interest object map was applied to the individual’s scans. An investigator analyzed the scans blindly to tasks associated with each scan. A base-line and average striatal C-raclopride binding potential was compared in the control condition and then in the task-related conditions and use of an ANOVA was determined and used to inform the mean of the RAC conditions.

Results and Performance

Participants in both groups scored well above 50% chance in the last 50 trials of the 400. Performance in both groups was great than chance FB thus indicating significant learning of the task. There was a linear trend reflecting the proportions correct across the linear trials and thus this also reflected a significantly better than average chance for all blocks of the trials. Participants in the FB group were able to learn the task after 100 trials after which performance stabilized. The RAC scans indicated the effect of the region (right and left: caudate vs. putamen vs. ventral striatum, i.e., 6 levels) was a significant difference in levels of DA across the regions that related to a performance control. The main effect group however or interaction of the regions were significant indicating the control condition was also comparable across both FB and PA groups and thus it was suitable as a baseline measurement of RAC. A significant interaction occurred between group, condition, and region that indicated a difference in RAC for active and control conditions and it differed greatly across regions between the FB and PA groups.

The mean percentage change in RAC was dramatically different between the right ventral striatum with left ventral striatum and this was statistically significant. All other comparisons were not significant. Within regions, no within-group differences were noted in the PA group. The right ventral striatal increased 5% in the PA during the task versus the control. The FB group displayed a marginally significant reduction in RAC in the right and left striatum when performing the active task compared to the control task (13% to the right and 6% reduction to the left). This indicated a release of dopamine during the FB task. One participant in the feedback group achieved a very low score during the feedback session of just above 50%. These participants scored were eliminated and following that, the FB group shows a significant reduction in RAC in the right ventral striatum when compared to the control task at a 17% reduction. This indicated a release of synaptic dopamine during the FB task. The percentage negative correlation between the mean percentages in the right ventral striatum indicated that the learning process was specifically related to the increase of dopamine in the right ventral striatum area.


Previous MRI studies have detected significant blood/oxygen levels that signal change in the striatum during a classification learning process, this study was the first to demonstrate the release of striatal dopamine during a corrective feedback version of the weather prediction therapy. The 13 – 17% reduction of RAC versus the control is consistent with the role of ventral striatal dopamine in mediating learning with feedback. It is consistent with and supported by a large body of literature in displaying the role of the ventral striatum is involved with learning tasks that involve probabilistic classification and how it’s activation is associated with feedback processing. It is thought to mediate probabilistic reverse learning which is affected and impaired by those with Parkinson’s disease while on medication. It was stated it was possible that due to fluctuating temporal activity, some “average” dopamine release was not detected. That said, the evidence does indicate that compared to healthy subjects, Parkinson’s Disease patients are impaired on the FB version of the therapy while when taken off medications, they are not affected. This suggests a dopamine ‘overdose’ and supports the proposal of tonic increase of dopamine with dopaminergic medications hides phasic changes in the dopamine release essential for learning. While there is a selective increase in dopamine release in the ventral striatum during FB based learning and not in PA based learning, FB task learning may potentially involve more effort and decision making and thus might explain this pattern. In conclusion, striatal dopamine is selectively released during FB based learning but not in PA probabilistic classification and the results have implications for understanding the mechanisms for feedback dependent and observational learning systems. It also helps to understand why patients with Parkinson’s disease have selective learning impairments on FB learning vs PA learning. This study also contributes to the explanation as to why there is a FB learning deficit in those on medication for Parkinson’s disease.

Author of the Original Study

Leonora Wilkinson — National Institute for Neurological Disorders and Stroke, Bethesda, Maryland and is a member of the Cognitive Motor Neuroscience Group, Sobell Department of Motor Neuroscience and Movement Disorders, Institute of Neurology, University College London, London, UK
Yen Foung Tai — Center for Neuroscience, Department of Medicine, Imperial College London, London, UK
Chia Shu Lin — Center for Neuroscience, Department of Medicine, Imperial College London, London, UK
David Albert Lagnado — Department of Psychology, University College, London UK
David James Brooks is also a member of the Center for Neuroscience, Department of Medicine, Imperial College London, London, UK as well as a member of the Institute of Clinical Medicine, Aarhus University, Aarhus, Denmark
Paola Piccini — is a member of the Center for Neuroscience, Department of Medicine, Imperial College London, London, UK
Marjan Jahanshahi — Cognitive Motor Neuroscience Group, Sobell Department of Motor Neuroscience and Movement Disorders, Institute of Neurology, University College London, London, UK

No items found.

Recent Posts