The Clock Drawing Test as a Dynamic Assessment Tool for Monitoring Cognitive Changes in Hospitalized Older Adults Across the Cognitive Spectrum, Including Dementia: A Prospective Observational Study

Introduction and Objective: Monitoring short-term changes in functional and neuromotor performance among older adults during a two-week hospital stay remains challenging in everyday geriatric practice. In this context, the present exploratory observational study aims at investigating the potential of the Clock Drawing Test, evaluated by means of quantitative geometric and kinematic parameters, as a method for capturing short-term fluctuations in visuomotor performance in routine multimodal inpatient care. The aim of the study is not to identify the isolated causal effects of single pharmacological or physical interventions, but to formulate hypotheses on the evolution of graphomotor trajectories within the framework of the usual clinical management. Methods: We conducted a prospective observational study involving 90 older adults hospitalized with chronic medical comorbidities and receiving multimodal inpatient care, including pharmacological treatments and physical rehabilitation techniques such as kinesiotherapy and TECAR therapy. Participants were classified into three diagnostic groups according to their cognitive status: no cognitive impairment (n = 40), mild cognitive impairment (n = 25), and dementia (n = 25). Results of the Clock Drawing Test performed at admission (T0) were compared with those performed at discharge (T14), focusing on contour integrity, numeral spacing, hand placement precision, and digital kinematic parameters. As the primary longitudinal outcome, paired T0 and T14 drawings were compared to evaluate short-term visuomotor changes. As a secondary exploratory analysis, baseline (T0) cognitive status classification was evaluated using Random Forest and Multilayer Perceptron models through participant-level 5-fold stratified cross-validation to determine how baseline neurocognitive and vascular profiles frame short-term graphomotor trajectories. Additionally, a binary (two-way) classification experiment (no cognitive impairment, n = 40, vs. cognitive impairment combining MCI and dementia, n = 50) was conducted, and the Structural Similarity Index (SSIM) was employed to objectively quantify whole-image morphological differences between CDT images obtained at admission (T0) and discharge (T14) using a standardized clock as ground truth. Results: Across the 90 complete pairs, significant Time × Regimen Cluster interactions were observed (p < 0.001). In Cluster 1 (multimodal physical and neurotrophic rehabilitation, n = 55), discharge evaluations showed significant improvements in contour closure gap (from 9.12 ± 2.45 mm to 1.78 ± 0.64 mm, p < 0.001), angular hand placement error (from 27.9° ± 6.8° to 4.1° ± 1.9°, p < 0.001), air-to-paper time ratio (from 1.84 ± 0.42 to 1.36 ± 0.31, p < 0.001), and Z-axis pressure stability (p < 0.001), whereas conceptual time-representation errors remained unchanged. Conversely, the small exploratory Cluster 2 subgroup that was treated with central nervous system depressants for acute inpatient insomnia or agitation showed a decline in graphomotor parameters (contour gap 2.15 ± 0.70 mm to 7.85 ± 2.30 mm, p = 0.016). This pattern could be a marker of either the pharmacodynamic sedation or the underlying acute neuropsychiatric conditions that warranted treatment. Conversely, Cluster 3 (standard medical maintenance, n = 27) remained stable (p > 0.20). For the secondary baseline classification task, the regularized Random Forest and Multilayer Perceptron models achieved promising internal cross-validation (out-of-fold) balanced accuracies of 86.00% and 88.13%, respectively. When reformulated as a two-way classification task (no impairment vs. MCI + dementia), out-of-fold overall accuracies increased to 92.22% (balanced accuracy 92.00%) for Random Forest and 94.44% (balanced accuracy 94.25%) for Multilayer Perceptron. Across 83 paired image sets, mean SSIM between T0 and T14 was 0.7494 ± 0.1016, while mean SSIM relative to the ground-truth clock changed from 0.7527 ± 0.0609 at T0 to 0.7352 ± 0.0687 at T14 (p < 0.001). Conclusions: Quantitative geometric and kinematic CDT parameters captured short-term changes in visuomotor execution over a 14-day hospitalization, while conceptual clock-drawing deficits remained associated with baseline diagnostic status in this exploratory observational cohort. Confounding by indication, practice effects and relief of pain or fatigue may also contribute to the observed trajectories in non-randomized multimodal care. Controlled trials and external multi-center validation cohorts are needed to confirm broader model generalizability and establish the independent effects of specific therapies.