How Can Clinicians Decide if AI-Enabled Risk Prediction Models Are Fit for Purpose for Real World Use?

Cognitive flexibility is a key component of cognitive control alongside top-down attention and inhibition, enabling us to quickly adapt behavior to an unpredictable and changing environment. This executive function varies widely between individuals and over the lifetime (). At its core, cognitive flexibility allows us to make decisions under uncertainty, based on cognitive rule representations whose selection depends on feedback from the environment. For example, identifying the right key of a set of unmarked keys to open a door requires remembering (i.e. using a mental map of) which key opens which door, but because such mental maps may be uncertain, a trial-and-error approach can augment the search until the right key is identified by the opening of the door. The internal representation of which key has to be used remains uncertain until the door finally opens. A fundamental question is how such cognitive flexibility for switching between cognitive rules (or keys) under uncertainty is achieved in the brain. The ability to adapt to changing task rules, even when they are not explicitly instructed, is called task switching and has often been studied with the Wisconsin Card Sorting Test (WCST) in humans (; ), monkeys (; ) and recurrent neural networks in computational work (). Performance in the WCST depends on multiple components, namely global and local switching, error monitoring and reward processing (; ; ), but only one study has tried to distil the impact of those sub-processes on interindividual performance variability in the WCST using a computational model that extends the traditional error type analysis of the WCST with subject-level model parameters of feedback processing (). In the brain, switching between cognitive representations such as task rules depends critically on activation of the prefrontal cortex (PFC) (), in line with the PFC’s general role in cognitive control tasks (; ; ), but also on activity at posterior parietal and occipital cortices (; ). Research in non-human animals has further shown that the mediodorsal thalamus and its interactions with the PFC are key to enabling cognitive flexibility (; ). This is especially true under task uncertainty (), where the informativeness of external cues depends on the uncertainty of the internal rule representation (  ). Accumulating evidence indicates that brain rhythms—neuronal oscillations—are fundamental for temporal coordination of neuronal processing in a variety of cognitive functions (; ; ; ). Two decades of work leveraging electroencephalography (EEG) to study overt task switching in relation to sensory representations have revealed that local theta (θ, 3–8 Hz) band oscillations in the PFC or the midline θ activity play a pivotal role in cognitive control (; ) including conflict monitoring (; ) and anticipation (; ). Importantly, θ oscillations implement cognitive control also in visual working memory (VWM) (; ), speaking for their general algorithmic role in implementing frontal control functions. Local field potential data from non-human primates further show that rhythmic θ-band activity in the anterior cingulate cortex (ACC) gives rise to prominent θ-oscillations during stimulus–response mapping (; ; ; ). In summary, θ oscillations originating in the PFC as well as ACC have been shown across various tasks depending on feedback-guided cognitive flexibility. In addition to the ubiquitous θ-band oscillations, there is EEG-backed evidence for the role of local gamma (γ, 30–90 Hz) (), parietal alpha (α, 8–12 Hz) (; ), and beta band (β, 13–30 Hz) (; ) amplitudes in task-switching although this evidence is less consistent. Again, in VWM research, local modulation of α oscillations has been established to support top-down or feed-back attentional selection (; ; ; ), whereas in contrast, γ oscillations and γ-range broad band power reflect bottom-up processing and broad band increases in neuronal firing rates, respectively (), carrying out complementary functions (). Crucially, the cross-frequency coupling of θ/α band oscillations with oscillations in higher β/γ frequencies via phase-amplitude coupling (PAC) enables the integration of top-down and bottom-up functions such that neurons coding for item-related information preferentially fire during particular θ phases () and that stimulus-related γ activity (>30 Hz) is locked to specific θ phases (; ). θ-γ PAC is therefore well-suited to serve as an integratory mechanism of slow and fast cognitive subprocesses. Accordingly, θ-γ PAC predicts performance variability in VWM (; ; ; ; ; ; ), attention (; ; ) and feedback-processing (; ), but its role in cognitive flexibility under uncertainty has been remained largely unexplored.