2.1.1. Design Goals

We wanted a corpus that i) resembles everyday speech, ii) embeds known AD-linked linguistic alterations in a controllable way, and iii) is fully reproducible. We therefore generated a balanced dataset of 440 transcripts (220 Control, 220 Early-AD-like), each comprising short paragraphs drawn from four neutral topics family, work, hobbies, and daily routine to avoid topical confounds.

2.1.2. Generative Process

Each transcript is composed of N_s sentences. For Controls, we sample:

while for AD-like we use μ=6 (slightly fewer sentences on average). A base sentence is drawn from a topic template and then noised according to class:

Controls (coherent baseline): grammatical sentence from a topic template; with small probability p f i l l e r c ≈ 0.1 a single filler (e.g., “um”, “you know”) is inserted; with p p a u s e c ≈ 0.5 a PAUSE token may appear.

AD-like (impaired speech profile):Shortening/reduced mean sentence length: we truncate at a target length L∼max(4, N (9, 3)).Fillers & pauses: insert with higher probabilities p f i l l e r A D ≈ 0.6 , p p a u s e A D ≈ 0.5 , at random positions to mimic hesitations.Semantic drift/vagueness: with P d r i f t ≈ 0.4 , append a vague follow-up clause (e.g., “the thing was there… it went where it goes”).Pronoun inflation: append simple connectors (and/then/so/but) followed by a pronoun (I/it/they…) 1 - 3 times to raise the pronoun ratio.Local repetition: allow repeated bi grams within the transcript to emulate perseveration.

All insertions are Bernoulli draws at random token positions; topics are randomly permuted across sentences to prevent the model from anchoring on a topic. Generation uses a fixed random seed to ensure exact reproducibility [22].

2.1.3. Rationale

These manipulations reflect documented early AD phenomena: shorter utterances, lexical impoverishment (lower diversity), higher pronoun use, disfluencies (fillers/pauses), and semantic drift. By controlling injection probabilities, we can later ablate which signals matter most. The dataset size of 440 transcripts was chosen to balance realism, interpretability, and statistical stability for a proof-of-concept study. This scale is sufficient to estimate logistic regression coefficients reliably in a high-dimensional but strongly regularized setting while remaining small enough to allow full control over the generative process. To assess robustness, we repeated training under varying class balances (60/40 and 70/30) and reduced dataset sizes (N = 300), observing stable AUC and AP values with only minor variance, indicating that results are not driven by a specific sample size or balance configuration.

We compute numeric descriptors that summarize lexical richness, fluency, and syntactic/semantic load from each transcript. Let the tokenized transcript have W words, S sentences, and F function words (per a standard stop word list). Fillers were counted using the explicit token list {“um”, “uh”, “erm”, “you know”}, pauses were encoded as the literal token “PAUSE”, and pronouns were defined as first- and third-person forms {“I”, “me”, “we”, “they”, “it”, “he”, “she”, “them”, “this”, “that”}. These lists are fixed and provided verbatim to ensure exact replication of feature calculations. Then:

Type-Token Ratio (TTR): TTR = | unique tokens W |

Lower values indicate reduced lexical diversity.

Repetition rate (bi-grams): build all bi-grams; if U is the set of unique bi-grams and R = {g ∈ U: count(g) > 1},

Higher values suggest perseveration.

Idea density (proxy): content words per 10 tokens,

Lower values indicate semantic impoverishment.

Mean sentence length: W S

Average word length: characters per token.

Pronoun ratio: ≠ pronouns W ;Content-word ratio: W − F W

Fillers per sentence and pauses per sentence: counts normalized by S.

Flesch Reading Ease (FRE) as a readability/complexity proxy:

where syllables are estimated via a rule-based heuristic.

Neurocognitive mapping. TTR, idea density, and sentence length relate to semantic memory and working memory; fillers/pauses index fluency and executive control; pronoun ratio reflects lexical retrieval difficulty.

We combine sparse lexical signals with dense biomarkers:

1) TF-IDF (1–2-grams, max 2000 features, min_df = 2).

Captures local phrases (e.g., “you know”, “I was”) and literal PAUSE.We use a token pattern that keeps single-character tokens, ensuring pronoun “I” is not dropped.

2) Engineered numeric features (above), scaled with MaxAbsScaler.

MaxAbs handles disparate scales while preserving sparsity when concatenated with TF–IDF.

A ColumnTransformer concatenates the two blocks safely, so fit happens only on training data inside the pipeline (prevents leakage).

We use Logistic Regression with L2 regularization (solver LBFGS, max_iter = 2000) for strong baseline performance and interpretability. The decision function is:

where positive coefficients increase odds of the AD class. This affords global explanations and straightforward clinical narratives.

Why logistic regression?

It is robust on small-to-moderate N with high-dimensional sparse inputs.Coefficients directly map to log-odds, enabling transparent biomarker stories.It sets a credible baseline before considering heavier models (SVMs, gradient boosting, transformers).

Stratified 75/25 train/test split preserves class balance.All transforms (vectorizer vocabulary, scaling) are learned within the Pipeline on the training fold only.No text length or topic metadata is fed as raw features to avoid trivial shortcuts.

(Optionally, one can add StratifiedKFold CV on the training set for model selection; we report final metrics on the held-out test set to approximate generalization.)

We assess complementary aspects of performance:

ROC AUC (threshold-free separability).Precision–Recall (AP) (robust to class imbalance; highlights positive-class retrieval).Confusion matrix at a default 0.5 threshold with precision/recall/F1 (from the classification report).Calibration curve (10 uniform bins) to examine probability reliability; optionally compute Brier score.Low-dimensional embedding (PCA 2D) of the fused feature space for qualitative separation.Global interpretability: sorted logistic coefficients for both AD-pushing and Control-pushing features.Caveat: coefficients can be affected by correlation among features; they are best complemented by permutation importance for stability checks.

Threshold selection (optional). If a specific clinical operating point is desired, we can select a threshold by Youden’s J or by cost-sensitive utility (e.g., higher recall at the expense of precision for screening).

To understand which signals drive performance:

Ablation-1 (Text-only): TF–IDF block alone → baseline AUC.Ablation-2 (Biomarkers-only): engineered numeric features alone.Full model: TF-IDF + biomarkers; compare deltas to quantify each block’s contribution.Feature group drops remove fillers/pauses features or pronoun-related features to test their marginal impact.Stability: repeat train/test splits with different seeds; compute 95% CIs via bootstrap (e.g., 1000 resamples of test scores).

Ablation experiments confirmed the complementary value of the two feature blocks. Using TF-IDF features alone yielded strong but reduced discrimination (AUC ≈ 0.92), while engineered linguistic biomarkers alone achieved moderate performance (AUC ≈ 0.85). The combined model substantially outperformed either block in isolation, demonstrating that lexical n-grams and cognitive–linguistic summaries capture distinct and additive information. These analyses guard against over-reliance on any single cue (e.g., the PAUSE token).

Pipeline: scikit-learn ColumnTransformer + Pipeline for atomic fit/transform/predict.Preprocessing: lowercasing; token pattern keeps single-character tokens; punctuation delimits sentences; PAUSE treated as a literal token to ensure the vectorizer “sees” it.Exports: CSV dataset, metrics, and 300-DPI figures (ROC, PR, confusion matrix, calibration, PCA, coefficient bar charts) are saved to a versioned folder and zipped for archival.Random state: fixed seed (42) for dataset generation and splitting; ensures exact reproducibility.

Although this study uses synthetic transcripts (no human subjects), the intended application involves patient speech. Key considerations for real deployments:

ASR variability: accents/noise introduce transcription errors; robustness should be tested with noisy ASR outputs [23]-[25].Demographic fairness: language usage varies by age, education, dialect models must be audited for subgroup performance [26]-[28].Clinical integration: screening tools should be assistive, not diagnostic; false positives/negatives must be communicated clearly.Data governance: privacy-preserving pipelines and consent procedures are mandatory when handling real audio/text [29].

The hybrid model that combined TF–IDF lexical features with engineered linguistic biomarkers achieved near-perfect classification performance on the synthetic dataset. On the held-out test set, the model reached a ROC AUC of 1.00 and an Average Precision (AP) of 1.00, suggesting that the simulated early AD signals were highly discriminative. To quantify statistical uncertainty, we computed 1000 bootstrap resamples of the held-out test set. The resulting 95% confidence intervals were AUC = 1.00 [0.98, 1.00] and AP = 1.00 [0.97, 1.00]. Although point estimates are perfect, the confidence intervals appropriately reflect finite-sample uncertainty and guard against over-interpretation. Figure 1Figure 1 shows the Receiver Operating Characteristic (ROC) curve. The curve follows the top-left corner of the plot with no deviation, demonstrating flawless sensitivity–specificity trade-off. This indicates that at nearly all thresholds, the classifier can separate early AD from control transcripts with zero overlap. Figure 2Figure 2 presents the Precision–Recall (PR) curve. Here, the model sustains perfect precision even as recall approaches 1.0, confirming that false positives were absent in the test set. Taken

Figure 1. ROC curve for early AD vs control classification (AUC = 1.00).

together, Figure 1Figure 1 and Figure 2Figure 2 confirm that the injected linguistic features reliably encode the class labels in this synthetic scenario.

At a conventional decision threshold of 0.5, the model achieved 100% sensitivity and 100% specificity (Figure 3Figure 3). All 55 early AD transcripts in the test set were correctly identified, and all 55 control transcripts were correctly classified. While such perfect results are unlikely in real-world data, they provide important proof that the engineered features capture known linguistic patterns of AD. Figure 4Figure 4

Figure 2. Precision-Recall curve for early AD vs control classification (AP = 1.00).

Figure 3. Confusion matrix at threshold = 0.5. Both classes are perfectly separated.

Figure 4. Calibration curve showing close alignment between predicted probabilities and observed frequencies.

further examine the calibration of predicted probabilities. The calibration curve closely follows the diagonal reference line, indicating that the model’s probability estimates are well-calibrated: a sample predicted with 70% probability of being AD is observed to belong to the AD class approximately 70% of the time. This is crucial for clinical decision support, as it suggests the probabilities may be interpreted as reliable risk estimates rather than arbitrary scores.

To gain qualitative insight into class separation, we visualized the fused feature space using t-SNE embedding (Figure 5Figure 5). Control and early AD transcripts formed two clearly distinguishable clusters, with minimal overlap. This separation confirms that the engineered biomarkers, combined with lexical n-grams, create a feature space that reflects the underlying linguistic differences simulated in the dataset. The visualization also supports clinical interpretability: early AD language (high pronoun use, fillers, pauses) maps into a distinct subspace compared to healthy controls (higher lexical diversity, longer sentences).

Interpretability analyses revealed which linguistic features most strongly influenced classification.

Features pushing toward early AD (positive coefficients) included:Pauses per sentence and fillers per sentence (strongest predictors).Pronoun ratio, consistent with AD patients substituting pronouns for specific nouns [30]-[33].

Figure 5. t-SNE embedding of fused feature representations. Early AD and control transcripts form separable clusters.

Lower syntactic complexity reflected by Flesch Reading Ease and shorter sentence length.Specific lexical tokens such as “um”, “uh”, “you know”, and vague referents like “it” [34].Features pushing toward controls (negative coefficients) included:Mean sentence length and content-word ratio, both indicative of richer, more complex speech.Idea density and average word length, proxies for semantic specificity.Content-heavy words tied to daily routines or work contexts (“sales”, “market”, “living room”).

Figure 6Figure 6 and Figure 7Figure 7 visualize these findings, showing the top features with the largest positive and negative log-odds coefficients, respectively. Figure 8Figure 8 further validates these results through permutation importance, which quantifies how much each feature contributes to model AUC when randomly shuffled. The overlap between coefficient-based and permutation-based rankings increases confidence in the stability of these predictors.

To illustrate how these biomarkers manifest in individual samples, Table 1 presents a subset of 20 transcripts with their computed linguistic features. Clear differences emerge between groups:

Early AD transcripts exhibit shorter mean sentence lengths, higher pronoun ratios, and significantly more fillers/pauses per sentence [35][36]. Their readability scores (Flesch Reading Ease) are higher, reflecting shorter, simpler utterances. Idea density is consistently lower.Control transcripts maintain longer sentences, higher lexical diversity (TTR), and greater content-word ratios, aligning with expected richer speech profiles [37]-[39].

This descriptive evidence corroborates the classifier’s feature importance findings and provides concrete examples of how subtle shifts in language structure can serve as early biomarkers of cognitive decline.

Figure 6. Top features contributing to early AD predictions (positive log-odds).

Figure 7. Top features contributing to control predictions (negative log-odds).

Figure 8. Permutation importance of top features, showing their effect on model AUC.

Table 1. Example subset of the synthetic speech dataset showing text, labels, and extracted linguistic biomarkers (20 samples).