This study introduces an interpretable modeling framework—CharMark—for identifying early linguistic biomarkers of dementia from transcribed speech. Our approach treats each transcript as a character sequence and models it using a first-order Markov Chain. From this, we compute a 27-dimensional vector representing the steady-state probability of each character (a–z and the space character). These vectors are then analyzed through clustering, dimensionality reduction, hypothesis testing, and logistic regression to evaluate group-level differences and the potential diagnostic value of the extracted features.

Figure 2 provides a conceptual overview of the CharMark pipeline.

We used transcripts from the DementiaBank Pitt Corpus, a widely cited dataset in dementia and speech-language research. All participants were asked to describe the Boston Cookie Theft picture, a standardized elicitation task designed to generate naturalistic but structured language samples [8].

The dataset included:

310 transcripts from 168 participants with Alzheimer’s Disease (AD)

Markov Chain modeling offers a mathematically grounded and interpretable framework for capturing sequential dependencies in language. Its ability to model transition dynamics with minimal assumptions makes it well suited for identifying subtle patterns that may serve as early digital biomarkers [9]. Building on our prior work using symbolic recurrence and character-level modeling to capture linguistic markers of cognitive decline [10, 11], we adopt a first-order Markov Chain approach in this study.

Each transcript was modeled as a first-order Markov Chain, where the probability of observing character sj depends only on the immediately preceding character si. Transition probabilities were estimated using smoothed frequency counts with Laplace smoothing (α=0.01), a small constant selected to mitigate zero-probability transitions while preserving the sparsity of natural language character distributions:Pij=nij+α∑u=1k(niu+α)where nij is the number of observed transitions from character si to sj, and k=27.

To assess robustness, we further conducted a sensitivity analysis varying α∈{0.001,0.005,0.01,0.05,0.1} while holding the classifier and splits fixed. Performance was stable across the range (AUC varied by ∼0.027, F1 by ∼0.035), confirming that α=0.01 is a reasonable and stable choice for all main analyses. Full numerical results are provided in Table 3 and trends are visualized in Figure 3.

To derive the long-run behavior of character usage, we computed the steady-state distribution π by solving the eigenvector equation:π=πPwith∑i=1kπi=1This steady-state vector, which reflects the long-term usage frequency of each character under the Markov process, was extracted for every transcript, forming a 552 × 27 feature matrix. These steady-state distributions serve as compact, interpretable fingerprints of linguistic structure across groups.

We applied k-means clustering to the steady-state vectors to explore group separability in an unsupervised setting. The optimal number of clusters was determined using silhouette analysis, which identified k=2 as the most stable solution corresponding closely to the ground truth diagnostic labels of healthy controls and dementia patients. This finding suggests that meaningful cognitive signals are embedded in the character-level structure of language, independent of any supervised learning.

To better understand the geometry of these features, we performed Principal Component Analysis (PCA) on the 27-dimensional steady-state vectors. PCA reduces dimensionality while preserving the directions of greatest variance, enabling clearer visual interpretation. As shown in Figure 4, the resulting projection reveals a striking separation between clusters, with transcripts from dementia subjects forming a more compact and shifted group compared to the wider distribution of controls. This emergent clustering reinforces the hypothesis that cognitive decline alters linguistic dynamics in ways that are both measurable and visually discernible at the character level.

To identify which characters contributed most to group-level differences, we conducted two-sample Kolmogorov-Smirnov (KS) tests on each of the 27 character distributions (dementia vs. control). The KS statistic quantifies the maximal distance between the empirical cumulative distribution functions:D=supx|Fdementia(x)−Fcontrol(x)|We selected the KS test due to its non-parametric nature and sensitivity to both location and shape differences between distributions, making it appropriate for features that may not follow Gaussian assumptions. Significance was evaluated at α=0.05 with Bonferroni correction to control for multiple comparisons across the full character set. This conservative adjustment reduces false positives, ensuring that only robust distributional shifts are flagged as candidate biomarkers.

To assess the predictive relevance of the steady-state features, we trained a Lasso Logistic Regression model using the 27 character probabilities. Lasso was chosen for its ability to perform both classification and feature selection, enabling a sparse and interpretable solution. We evaluated model performance using the area under the Receiver Operating Characteristic curve (ROC-AUC).

To ensure comparability, we used stratified 5-fold cross-validation and reused the same splits across CharMark and all baselines (TF–IDF + Logistic Regression, BERT-base). Where participant identifiers were available, all transcripts from a participant were confined to a single fold. We report mean ± SD across folds for AUC, F1, precision, recall (sensitivity), specificity, and accuracy.

To prevent data leakage across folds, all transcripts from a given participant were confined to a single fold. This ensured that no subject contributed data to both training and testing sets. Participant identifiers (Pitt Corpus IDs) were used to enforce group-level assignment when constructing cross-validation splits.

To evaluate probability calibration, we computed Brier scores and Expected Calibration Error (ECE) across cross-validation folds, and plotted reliability curves. Threshold sensitivity was assessed by comparing default classification at p=0.5 with thresholds optimized for Youden’s J statistic and for maximal F1.

While many dementia studies utilize acoustic features, we focused on text-based analysis to enhance interpretability, reproducibility, and cross-linguistic applicability. Character-level modeling captures fine-grained language disruptions such as excessive pausing, rigid phrasing, and letter-specific anomalies that are often masked in higher-level or audio-driven analyses. Moreover, this approach is computationally efficient, robust to noise, and aligned with emerging needs for lightweight, privacy-preserving digital biomarkers in remote or resource-constrained environments.

TF–IDF + Logistic regression: We extracted word unigrams and bigrams with min_df = 5 and max_df = 0.8 and trained an L1-penalized logistic regression model (C = 1.0, class-balanced, max_iter = 1,000).

BERT-base (uncased): We fine-tuned bert-base-uncased for sequence classification (maximum length 128, batch size 16, learning rate 2×10−5, 2 epochs) with simple per-fold oversampling on the training split to balance classes.

For both baselines we reused the same stratified 5-fold cross-validation splits as for CharMark to ensure an apples-to-apples comparison. We report mean ± SD across folds for AUC, F1, precision, recall, specificity, and accuracy.

To further contextualize CharMark’s contribution, we evaluated two ablations. First, a frequency-only baseline using raw 27-dimensional character frequency vectors. Second, a space-only baseline using the proportion of whitespace tokens in each transcript. Both were trained with the same L1-penalized logistic regression classifier under identical 5-fold splits.

We first examined steady-state character distributions to identify linguistic signals that differed significantly between groups. The Kolmogorov-Smirnov (KS) test revealed that several characters, including the space character “ ,” “n,” and “i,” showed statistically significant distributional shifts between dementia and control groups (Bonferroni-corrected p<0.05). These characters reflect changes in pacing, repetition, and lexical structure often observed in cognitive decline. Table 1 summarizes the top-ranked features based on KS statistic.

Among all features, the space character emerged as the most distinctive. Figure 5 visualizes the empirical distribution of steady-state space probabilities across groups. Dementia transcripts exhibit a rightward shift in the distribution, reflecting longer or more frequent pauses; likely tied to disrupted fluency. Reflecting greater whitespace token frequency, which may indicate disrupted fluency at the textual level.

To further probe this finding, we plotted a rolling mean of the space character probability across transcript indices (Figure 6). While average values remain elevated in dementia subjects, healthy controls display greater variance, potentially reflecting more adaptive or dynamic speech rhythm. The convergence of elevated mean and lower variance in dementia aligns with prior hypotheses about pacing rigidity in cognitive decline.

To evaluate the diagnostic potential of our extracted features, we trained a Lasso Logistic Regression model using the 27-dimensional steady-state character vectors. The model was trained in a binary classification setting (dementia vs. control), and the resulting Receiver Operating Characteristic (ROC) curve is shown in Figure 7. The model achieved an area under the curve (AUC) of 0.806, indicating strong discriminative performance from the character-level features alone.

Lasso regularization yielded a sparse solution, emphasizing only the most informative features. Notably, the space character remained among the most predictive features selected by the model, reinforcing its potential role as a candidate digital biomarker. To contextualize performance, TF–IDF + logistic regression achieved AUC 0.801±0.023 and BERT-base achieved AUC 0.8165±0.015 on the same 5-fold splits. Alongside TF–IDF and BERT baselines, both ablations were substantially weaker. The frequency-only model achieved AUC 0.782±0.024 and F1 0.701±0.026, while the space-only model dropped further to AUC 0.695±0.048 and F1 0.612±0.052. These results confirm that CharMark’s lift derives from its transition dynamics, not marginal frequencies or space proportion. Full metrics for all models appear in Table 2.

CharMark performance was stable across the tested range of α. AUC varied by approximately 0.027 and F1 by 0.035, with Precision, Recall, Specificity, and Accuracy similarly consistent. We therefore retain α=0.01 for all main analyses. The trends are visualized in Figure 3, and detailed numerical results are provided in Table 3.

CharMark demonstrated good discrimination (AUC = 0.803 ± 0.029) with moderate calibration error (Brier = 0.183 ± 0.013, ECE = 0.104 ± 0.022). At the default threshold (p=0.5), F1 = 0.719 ± 0.025, Precision = 0.672 ± 0.028, Recall = 0.699 ± 0.047, and Accuracy = 0.734 ± 0.020. Threshold optimization via Youden’s J and F1-maximization yielded only minor changes, indicating that the model is reasonably well calibrated for screening applications. Reliability curves are shown in Figure 8.

We further performed a second-order Markov ablation, in which transcripts were modeled as bigram-to-bigram transitions over 272=729 states. Despite the expanded state space, performance did not improve: AUC = 0.795 ± 0.026 and F1 = 0.707 ± 0.025, both slightly below the first-order model. This supports the choice of first-order modeling in the present study, balancing statistical stability and interpretability.

While scalar features like steady-state probabilities offer useful diagnostic signals, structural properties of the underlying transition networks also carry valuable insights into how linguistic rigidity manifests. To qualitatively explore this, we visualized first-order Markov transition networks from two representative transcripts, one from a dementia subject and one from a healthy control (Figure 9).

The dementia network is dominated by self-loops and short-range transitions among a limited subset of characters, suggesting repetitive or constrained lexical production. In contrast, the control network exhibits greater transition diversity and broader edge connectivity, consistent with richer, more adaptive language use. These structural differences offer a complementary lens into how cognitive impairment alters the foundational dynamics of speech.

Group-level demographics are summarized in Table 4 (Number of participants,gender, sex, age, education, MMSE).