NeuroDx-LM: EEG Diagnostic Adaptation

• NeuroDx-LM is an EEG-based clinical large-scale model that adapts a pre-trained backbone through a diagnosis-oriented framework to detect neurological disorders.
• It introduces a Selective Temporal-Frequency Embedding mechanism and a two-stage Progressive Feature-Aware Training strategy to refine temporal and spectral features.
• The model uses LoRA-based parameter-efficient tuning to handle heterogeneous EEG data, achieving improved accuracy in seizure and schizophrenia detection benchmarks.

NeuroDx-LM is an EEG-based clinical large-scale model for automatic neurological disorder detection. In the paper that introduces it, the model is instantiated for seizure detection on CHB-MIT and schizophrenia detection on a schizophrenia EEG dataset, and is presented not as a foundation model trained entirely from scratch, but as a diagnosis-oriented adaptation framework built on top of a pre-trained EEG large-scale model, specifically LaBraM-Base. Its two defining contributions are a Selective Temporal-Frequency Embedding (STFE) mechanism and a Progressive Feature-Aware Training strategy, with the stated goal of improving clinical EEG diagnosis under low signal-to-noise ratio, heterogeneous disease manifestations, and limited labeled data (Jin et al., 11 Aug 2025).

1. Clinical setting and conceptual scope

The NeuroDx-LM paper situates EEG as central to diagnosing disorders such as epilepsy and schizophrenia, while emphasizing that manual reading by neurologists is time-consuming. It argues that existing EEG models often generalize poorly because EEG recordings vary across devices, channels, patients, and disease manifestations, and that even recent EEG pre-trained models face weakened transfer to disease detection because of the low signal-to-noise ratio of EEG, the multi-faceted nature of neurological abnormalities, and the scarcity of labeled disease-specific EEG data. The paper also identifies a practical systems issue: standard full-parameter fine-tuning of large models is expensive in GPU memory, which makes deployment harder (Jin et al., 11 Aug 2025).

Within that framing, NeuroDx-LM is defined as a clinical large-scale model for EEG-based neurological disorder detection, but the “large-scale” designation comes from its use of a pre-trained EEG large-scale backbone rather than from training a giant new model from scratch. The contribution lies in diagnosis-oriented adaptation: the model first learns normal/abnormal discrimination from a larger annotated EEG corpus and then specializes to target disorders using both temporal and frequency information. The paper’s central claim is therefore architectural and curricular rather than scale-driven.

2. Architecture and signal representation

At a high level, NeuroDx-LM consists of preprocessed raw EEG input, patching into 1-second subsegments, the Selective Temporal-Frequency Embedding front end, a pre-trained EEG backbone (LaBraM-Base), and a downstream classifier. All EEG recordings are segmented into non-overlapping 10-second segments, and each 10-second segment is then further divided into 10 non-overlapping 1-second patches. The paper denotes each EEG patch as

$x_{i,j} \in \mathbb{R}^{C \times A \times T},$

while also noting that this notation is somewhat inconsistent because $x_{i,j}$ is described as a patch but includes $A$, the number of patches. After STFE processing, the model forms a sequence

$$\{z_1, z_2, \dots, z_M\}, \quad z_i \in \mathbb{R}^d,$$

prepends a classification token,

$[\langle cls \rangle, z_1, z_2, \dots, z_M],$

and passes the sequence into LaBraM-Base (Jin et al., 11 Aug 2025).

STFE is the model’s main architectural novelty. Its purpose is to combine information from the time domain and the frequency domain while allowing the training procedure to control how much spectral information is used at different stages. The temporal branch performs multi-scale temporal analysis using three independent temporal convolutional blocks, each containing a convolution, GELU activation, and Group Normalization, with kernel sizes

$\{15, 3, 3\}.$

The paper defines the temporal block as

$$z^{\text{time}'} = \text{GN}(\text{GELU}(\text{Conv}(x(t)))).$$

The frequency branch first applies FFT, retains the amplitude spectrum, and passes it through four convolutional blocks, again with convolution, GELU, and GroupNorm, using kernel sizes

$\{3, 15, 3, 3\}.$

The intended meaning of the paper’s frequency-branch expression is

$$z^{\text{freq}'} = \operatorname{GN}\!\left( \operatorname{GELU}\!\left( \operatorname{Conv}\!\left( \left| \mathcal{F}\{x(t)\} \right| \right) \right) \right),$$

where $\mathcal{F}\{\cdot\}$ is the Fourier transform implemented by FFT and $x_{i,j}$0 is the amplitude spectrum.

Temporal and frequency features are then fused with a gating variable $x_{i,j}$1, which scales the contribution of the frequency branch. In the experiments, this gate is not learned as a continuously varying controller; it is used as a stage-dependent switch:

• Stage 1: $x_{i,j}$2
• Stage 2: $x_{i,j}$3

The model size reported in the comparison table is 5.8M, the same as LaBraM. The paper therefore characterizes NeuroDx-LM as a clinically adapted EEG large-scale model, but in parameter count it remains modest by language-model standards (Jin et al., 11 Aug 2025).

3. Progressive Feature-Aware Training and LoRA adaptation

The second main novelty is the Progressive Feature-Aware Training strategy, described as a two-stage process inspired by curriculum learning. The idea is to progress from easier and broader abnormality discrimination to narrower disease-specific diagnosis. In Stage 1, the model is trained on the Temple University Abnormal EEG Corpus (TUAB) using normal vs abnormal labels. In this stage, the paper freezes the STFE module and the main model parameters, inserts LoRA modules, and disables the frequency branch by setting $x_{i,j}$4. In Stage 2, the model is fine-tuned on a smaller disease-specific dataset such as CHB-MIT or the schizophrenia dataset, activates the frequency branch with $x_{i,j}$5, merges the Stage 1 LoRA into the base model, and introduces a new disease-specific LoRA module (Jin et al., 11 Aug 2025).

For a pre-trained weight matrix

$x_{i,j}$6

LoRA introduces a low-rank update

$x_{i,j}$7

with

$x_{i,j}$8

The adapted forward pass is written as

$x_{i,j}$9

Only $A$0 and $A$1 are updated; $A$2 remains frozen.

Stage	Data	Key setting
Stage 1	TUAB	normal/abnormal discrimination; $A$3; LoRA inserted
Stage 2	CHB-MIT or schizophrenia dataset	disease-specific detection; $A$4; Stage 1 LoRA merged; new disease-specific LoRA

The implementation details reported in the paper are: backbone LaBraM-Base, batch size 64, AdamW, layer decay and warm-up, Stage 1 learning rate $A$5 for 1 epoch, Stage 2 learning rate $A$6 for 20 epochs, LoRA rank $A$7, and LoRA scaling $A$8. Hardware details reported from the ablation section are CUDA 11.8 and an RTX 3090 GPU. The paper does not specify the exact learning rate schedule shape, weight decay value, dropout, hidden dimension $A$9, classification head details, inference speed, random seeds, or number of runs (Jin et al., 11 Aug 2025).

4. Datasets, preprocessing, and quantitative performance

NeuroDx-LM uses three datasets. TUAB is used only in Stage 1 and is described as 23-channel EEG, sampling rate 256 Hz, total duration 1169 hours, with normal vs abnormal labels. CHB-MIT includes 22 pediatric epilepsy patients, 961 hours of scalp EEG, 198 seizure events, original sampling rate 256 Hz, with seizure onsets labeled by experienced neurologists. The schizophrenia dataset includes 14 schizophrenia patients, 14 healthy controls, 19 electrodes in the international 10–20 system, and original sampling rate 250 Hz. The common preprocessing pipeline removes channels not conforming to the international 10–20 system, applies a 0.1–75 Hz bandpass filter, applies a 50 Hz notch filter, resamples all EEG to 200 Hz, and segments signals into non-overlapping 10-second segments (Jin et al., 11 Aug 2025).

On CHB-MIT, NeuroDx-LM achieves:

• Accuracy: 0.7903
• ROC-AUC: 0.5903
• PR-AUC: 0.2617

On the schizophrenia dataset, NeuroDx-LM achieves:

• Accuracy: 0.8246
• ROC-AUC: 0.8965
• PR-AUC: 0.9114

The paper compares NeuroDx-LM against SPaRCNet, CNN-Transformer, ST-Transformer, BIOT, EEGPT, and LaBraM. For schizophrenia, the comparison to LaBraM is explicit: Accuracy improves from 0.7587 to 0.8246, ROC-AUC improves from 0.8668 to 0.8965, and PR-AUC improves from 0.8991 to 0.9114. On CHB-MIT, the gains are narrower: Accuracy is 0.7903, narrowly above EEGPT (0.7902) and above LaBraM (0.7550), while PR-AUC 0.2617 is the best in the table (Jin et al., 11 Aug 2025).

The ablation study isolates three factors: embedding choice, whether Stage 1 is used, and how Stage 2 is performed. The best configuration is STFE, Stage 1: Yes, Stage 2: Addition of new LoRA, with:

• CHB-MIT: 0.7903 / 0.5903 / 0.2617
• Schizophrenia: 0.8247 / 0.8965 / 0.9114

Compared with TE-only under the same training pattern, performance drops to:

• CHB-MIT TE-only: 0.7526 / 0.6027 / 0.2501
• Schizophrenia TE-only: 0.6285 / 0.7485 / 0.7651

Skipping Stage 1 hurts performance, especially on schizophrenia. With STFE and Stage 2 addition but no Stage 1, the results drop to:

• CHB-MIT: 0.7902 / 0.5241 / 0.2165
• Schizophrenia: 0.7344 / 0.8536 / 0.8805

If Stage 2 simply reuses Stage 1 LoRA rather than adding a new disease-specific LoRA, performance also declines:

• CHB-MIT: 0.7850 / 0.5557 / 0.2314
• Schizophrenia: 0.7569 / 0.8534 / 0.8556

The paper additionally states that using LoRA reduces GPU memory usage by 7% compared with full-parameter fine-tuning (Jin et al., 11 Aug 2025).

5. Interpretation, limitations, and practical significance

The paper presents NeuroDx-LM as neurophysiologically plausible. It argues that seizure EEG often contains abnormal transient waveforms, synchronization events, and rhythmic discharges, while schizophrenia-related EEG differences are often subtler and can involve altered spectral organization across bands and network rhythms. Within that interpretation, the Stage 1 emphasis on temporal features is aligned with spikes and sharp waves, while the Stage 2 addition of frequency features is aligned with disease-specific rhythmic structure. The model’s empirical profile is consistent with this design: STFE outperforms TE alone and TE+FE without the selective progression mechanism, and Stage 1 is particularly valuable when the disease-specific dataset is small (Jin et al., 11 Aug 2025).

At the same time, the paper is explicit or implicit about several limitations. Generalization across disorders is not fully demonstrated, because only seizure and schizophrenia detection are tested. Clinical deployment details such as real-time inference, robustness to artifacts, and cross-hospital transfer are not evaluated. Data split details are underspecified. Architecture details of the embedding fusion and classifier head are incomplete in the manuscript, and some equations contain formatting errors. The paper also notes that although the introduction claims improved interpretability, there are no saliency maps, attention visualizations, embedding projections, or frequency-band importance plots. Finally, the “dynamic gating” $$\{z_1, z_2, \dots, z_M\}, \quad z_i \in \mathbb{R}^d,$$0 is, in practice, a stage-specific binary switch rather than a learned adaptive controller, at least as described in the experiments (Jin et al., 11 Aug 2025).

In practical terms, NeuroDx-LM is best understood as a diagnosis-oriented adaptation framework for EEG large models. Its significance lies in combining a temporal-frequency embedding front end, a curriculum-like two-stage adaptation process, and LoRA-based parameter-efficient tuning. The paper’s own practical takeaway is that the method is especially worth considering when a practitioner has access to a pre-trained EEG backbone, a relatively large generic normal/abnormal EEG corpus, only a small labeled dataset for a specific disorder, and a need to incorporate both temporal and spectral EEG signatures.

6. NeuroDx-LM as a broader neurodiagnostic modeling pattern

Subsequent and adjacent work has used “NeuroDx-LM” less as the name of one EEG classifier and more as a broader framing for diagnosis-oriented neurological language and multimodal systems. The clearest example is NeuroLex, which is described by its authors as a partial but meaningful instantiation of a “NeuroDx-LM”-style system: it is not a full multimodal neurodiagnostic foundation model, but it is explicitly designed as an EEG-specialized language backbone for report understanding and generation, and as a future decoder backbone for multimodal EEG–language systems. NeuroLex is trained purely on EEG report text, uses an EEG-specific tokenizer, and reports Avg F1 0.575 on information extraction, ROUGE-L 0.707, BERTScore 0.956, Fact-F1 0.941 on summarization, and Neg-Adv F1 0.683 with Contr-Rate 0.058 in robustness tests. The paper’s own conclusion is that NeuroLex is best seen as the textual core or decoder of a NeuroDx-LM, rather than a complete neurodiagnostic LLM in the full multimodal sense (Yin et al., 17 Nov 2025).

A related neuroimaging direction is the modular MRI → report → LLM reasoning → diagnosis pipeline for neurodegenerative dementia differential diagnosis. That framework converts 3D T1-weighted brain MRI into a synthetic radiology report and then uses GRPO-tuned 8B LLMs for differential reasoning. Its best model, LLaMA-3.1-Instruct GRPO 8B, reaches BACC 67.33 and M-F1 65.09, slightly above the paper’s 3D Vision Transformer and SVM atrophies baselines. This suggests a NeuroDx-LM-style system can be organized around explicit textualized neuroimaging findings rather than only direct opaque classification, although the paper also notes the absence of external validation, calibration, and explicit faithfulness metrics (Zamai et al., 26 May 2025).

Benchmarking work has also clarified the limits of multimodal neurodiagnostic LLMs. NeuroVLM-Bench evaluates 20 frontier multimodal models on 2D brain MRI and CT with structured outputs including diagnosis, subtype, modality, sequence, plane, and confidence. The benchmark finds that modality and plane prediction are nearly solved, while diagnostic reasoning—especially detailed diagnosis / subtype—remains difficult. On the held-out zero-shot test, Gemini 2.5 Pro achieves diagnosis Macro-F1 0.576743 but only detailed diagnosis Macro-F1 0.322712. The benchmark therefore supports a more limited operational view of any future NeuroDx-LM: strong technical attribute inference, moderate coarse diagnosis, weak fine-grained subtype reasoning, and a need for explicit evaluation of calibration, abstention, structured-output validity, and cost (Dineva et al., 25 Mar 2026).

On the report-understanding side, Neu-RadBERT shows that neuro-specific report classification can be extremely strong when the task is narrow and the model is an encoder rather than an autoregressive generator. On brain CT and MRI reports from mechanically ventilated ARF patients, Neu-RadBERT with MLM pretraining and oversampling reaches up to 98.0% accuracy for acute brain injuries, whereas Llama-2-13B peaks at 67.5% binary classification accuracy on the acute-versus-not-acute task. This provides complementary evidence that a broader NeuroDx-LM ecosystem may need both encoder-classifier components for high-throughput diagnostic extraction and generative components for explanation or report generation (Singh et al., 1 Oct 2025).

Finally, work on RE-MCDF and BrainNet-MoE points toward two further extensions of the NeuroDx-LM idea. RE-MCDF introduces a generation–verification–revision closed-loop architecture for neurology EMR diagnosis and reaches F1 44.11 on NEEMRs and 40.48 on XMEMRs with Qwen2.5-7B-Instruct, emphasizing multi-expert reasoning, evidence prioritization, and explicit inter-disease logical constraints (Shen et al., 1 Feb 2026). BrainNet-MoE reaches ACC 82.76 and F1 80.07 for AD, LBD, and cognitively normal controls from structural connectivity, using disease-specific expert groups and a disease gate over brain sub-networks (Zhang et al., 5 Mar 2025). Taken together, these systems suggest that “NeuroDx-LM” increasingly denotes a broader research agenda: neurologically specialized models that combine modality-specific structure, disease-aware routing, and clinically meaningful reasoning, rather than a single monolithic model family.