Why saliency maps are supposed to build trust in the first place
A deep learning model that reads EEG and outputs "mild cognitive impairment" is a black box: the clinician sees a label, not a reason. Explainability methods try to open the box. Saliency maps highlight which parts of the input — here, which time segments or electrode channels — most influenced the model's decision [2][4]. Attention mechanisms do something similar by weighting which parts of the signal the model leaned on [1]. SHAP assigns each input feature a share of the blame or credit for the output [2].
The logic behind using them clinically is intuitive: if a model flags a patient and the saliency map points to the same brain regions a neurologist would find suspicious, the flag feels checkable rather than arbitrary. The survey of deep learning for dementia detection frames saliency maps and SHAP explicitly as tools "to foster clinical trust" [2], and the cognitive-decline prediction work describes making the model understandable to humans as a way to "foster trust and ensure responsible AI practices" [1]. That is a stated design goal, not a measured outcome.
What the EEG studies actually measured — accuracy, not trust
The diagnostic studies deliver on the model side. One framework detects Alzheimer's and Creutzfeldt–Jakob disease from EEG and produces spatio-temporal saliency maps for every EEG segment, giving a view of which channels and moments drove each call [4]. Another uses a deep generative model, Diff-EEG, to augment EEG data and simulate disease phases from mild cognitive impairment onward, with saliency maps highlighting critical brain regions [3]. The cognitive-decline work uses attention to show which EEG frequencies, including high frequencies, mattered in early deficits [1].
In each case the explainability output is evaluated as a property of the model — does it localize plausibly, does it point at meaningful regions — not as an intervention tested on people. None of these studies reports asking clinicians whether they trusted the model more, whether they changed a diagnosis, or whether they accepted the tool in practice [1][3][4]. The survey likewise covers explainable AI techniques for dementia detection without reporting clinician acceptance data [2]. So the honest reading is that the field has built the trust-building machinery and asserted its purpose, but has not yet put the assertion to a clinical test in these EEG studies.
The one piece of evidence that points toward trust
The closest thing to direct evidence comes from a broader clinical-analytics project using EEG to track early cognitive decline. There, the author reports that model interpretation "not only increases clinicians' trust in the AI" but serves a further purpose in the clinical workflow [5]. That is a claim about interpretation raising trust, made in the context of EEG-based cognitive monitoring — which is exactly the setting in question.
It is worth being precise about how much weight this carries. It comes from a single thesis-level source, and the trust claim is reported rather than demonstrated through a controlled comparison of clinicians using explained versus unexplained model output. It is consistent with the design rationale in the diagnostic papers [1][2], but consistency is not the same as confirmation. The reasonable summary is that interpretation plausibly helps clinician trust, with one source asserting it does, while the diagnostic EEG literature has not yet tested it.
What would actually settle the question
To know whether saliency or graph attributions improve clinician trust, you need a study where clinicians use the model with and without the explanation and you measure the difference — in trust ratings, in agreement with the model, in diagnostic accuracy, or in how often they override it. The current EEG work stops one step short: it produces explanations and argues they should help [1][2][4], and one clinical-analytics source reports that they do [5].
There is also a design caution buried in the diagnostic work. One framework notes that future iterations should provide further explanation specifically "to enhance trust" [4] — an admission that the current explanation is not yet considered sufficient. That suggests the question is not simply whether explanations help, but whether a given explanation is legible enough to a clinician to be useful. A saliency map that highlights a plausible brain region may reassure; one that highlights noise may do the opposite, and nothing in these studies tells us which happens with clinicians in the loop.
Sources used in this answer
Exploring the potential of explainable deep learning for EEG-based cognitive decline prediction
Presents a deep learning approach for predicting cognitive decline from EEG with attention-based explainability, but does not address clinician trust or acceptance.
A Survey on Deep Learning Techniques for Dementia Detection Using EEG Signals
Surveys deep learning for EEG-based dementia detection and highlights explainable AI methods such as SHAP and saliency maps as means to foster clinical trust, without reporting clinician trust or acceptance data.
Deep Generative Models for Alzheimer's Disease Diagnosis
Uses a deep generative model (Diff-EEG) for EEG data augmentation in Alzheimer's diagnosis, with saliency maps highlighting critical brain regions, but does not discuss clinician trust.
EEGDecoder-x: an explainable deep learning framework for cross-subject EEG-based detection of Alzheimer's and Creutzfeldt–Jakob disease
Introduces an explainable framework detecting Alzheimer's and Creutzfeldt–Jakob disease from EEG with spatio-temporal saliency maps, noting that future iterations should provide further explanation to enhance trust.
Interpretable Analytics for Clinical Applications
Reports that model interpretation increases clinicians' trust in AI in the context of EEG-based tracking of early cognitive decline.