#6 — At the Intersections — July 4, 2026

What happens when the system designed to train professionals is the very thing that holds them back? That's one of the boundaries we're exploring on ComplexityPod, where we look at how systems shape different domains of research. Here is this week's issue. This week's issue is focused on the boundaries of systems. And we're looking at that at very different scales. First, the environmental conditions affecting professional training, then the neural coordination between two collaborating people. From there, we turn to the process of classification. How machine learning uses color to identify objects in space, and how speech patterns might map to mental states. And the issue ends by considering conclusions that are formed without a foundation of formal data. So, starting with postgraduate medical education. The results of supervision practices like feedback and observation can be inconsistent. And this work, with member Sayra M Cristancho, applies a systems perspective to that inconsistency. It argues the effectiveness of any practice is shaped by the conditions of clinical work—workload, team structure, local culture. These things create patterns that can either help or hurt the learning process. So if a trainee's progress stalls, the problem might not be with the feedback itself, but with the larger system. Exactly. Feedback is only effective if a learner gets a chance to apply it. Observation is more useful if it's part of a continuous relationship. The research suggests that improving training isn't just about individual practices. You have to improve the conditions that allow those practices to connect and support learning. Otherwise, you're just asking people to optimize their actions within a system that limits what they can achieve. So from that large clinical system, we move to a much smaller one: the dyad. A study with members including Lena Palaniyappan examines the brain mechanisms that support coordinated action between two partners. They used dual-brain EEG hyperscanning on pairs of people assigned leader and follower roles. They were measuring both brains at the same time while they performed tasks. One task was mirroring a partner's movement, the other was a more complex transformation of that movement. Behaviorally, mirroring was faster, but only for the followers. And at the neural level? During the harder task, there was sustained frontoparietal engagement for everyone. But for the simpler mirroring task, the brain activity was role-specific. Leaders showed a kind of anticipatory gating, while followers had a post-response inhibitory rebound. The analysis pointed to a leader-specific predictive signal and a follower-specific adaptive signal. So coordination isn't a single process. It's a role-asymmetric architecture, where top-down prediction and bottom-up adaptation are functionally distinct. And identifying these role-specific signatures could be useful. They might serve as biomarkers for clinical populations where interpersonal coordination is disrupted, like in schizophrenia and autism spectrum disorder. Next, we shift to classification in astronomy. Specifically, how to identify globular clusters in large sky surveys. They're useful for understanding a galaxy's history, but they're hard to find because of contamination from other objects like stars and background galaxies. This research, with member Pauline Barmby, looks at how to select these clusters using only color data, in preparation for upcoming surveys. They tested different machine learning classifiers on archival data. And what's the best that machine learning can do with just color? How clean is the selection? Using all fifteen available colors resulted in a minimum contamination rate of about thirty percent. So even in the best case, almost a third of what you identify isn't a globular cluster. And using a more efficient method—just the first four principal components of the colors—achieves the same result. No better, no worse. Right. It's an improvement over traditional methods, which could double that contamination rate, but it shows the limits of using only color. The work concludes that to reduce contamination further, that photometric information has to be supplemented with other data, like morphology from space missions or near-infrared photometry. From classifying stars to classifying mental states. A systematic evaluation, also with Lena Palaniyappan, looks at the evidence connecting speech-based markers to mood states in Bipolar Disorder. The goal is to predict the emergence of the disorder non-invasively, using markers that can be gathered at a large scale. The analysis looked at 43 studies to see where the evidence stands. It found an emerging focus on mapping mood states to automated speech features. Speech could distinguish Bipolar Disorder from schizophrenia and depression. For example, manic states were linked to pressured speech, while depressive states showed more personal pronouns and less verbal fluency. But that's just a correlation. The review found that attempts to replicate these observations were limited. So the conclusion is that the evidence is currently insufficient for clinical use. It can't be used reliably for diagnosis or to predict a relapse. The recommendation is for future work to use natural language processing in longitudinal and cross-linguistic studies to build a stronger evidence base. And that brings us to the final piece from Sayra M Cristancho, called "When I say … lore." It’s a reflection on the narratives that are formed in the absence of formal data. And the contribution notes that no datasets were generated or analyzed for the work itself, which underscores the point. That concludes this issue of At the Intersections. We'll be back next week with more research summaries. From ComplexityPod, thank you for listening.