Malaria entomological surveillance in Kinshasa Province: A longitudinal study.

Parkinson’s disease (PD) is a chronic, progressive neurodegenerative disorder that affects motor control, speech, and quality of life for millions of individuals worldwide [–]. As the disease advances, patients experience gradually worsening motor symptoms such as bradykinesia, rigidity, tremor, and speech impairment [,]. Accurate assessment of symptom severity is therefore central to clinical management, treatment adjustment, and the evaluation of disease progression [,]. The Unified Parkinson’s Disease Rating Scale (UPDRS) remains the most widely used clinical instrument for quantifying PD severity [–]. Despite its clinical acceptance, UPDRS assessment has several well-recognized limitations. First, it requires the physical presence of the patient in a clinical setting, which can be burdensome for individuals with mobility impairments and costly for healthcare systems. Second, assessments are typically performed at infrequent intervals, often every several months, providing only sparse snapshots of a disease process that evolves continuously and exhibits substantial short-term fluctuations. Third, UPDRS scoring is inherently subjective, relying on expert judgment and clinical experience, which introduces inter-rater variability and limits reproducibility across settings and practitioners [,]. These limitations have motivated growing interest in telemonitoring approaches that enable remote, frequent, and objective assessment of PD symptoms. Advances in digital health technologies have facilitated collecting patient data in non-clinical environments, reducing logistical barriers while increasing temporal resolution [,]. Among the available modalities, speech recordings are particularly attractive: speech production is strongly affected by motor impairment in PD, voice acquisition is non-invasive, and sustained phonations can be reliably collected using simple, self-administered protocols [–]. Prior work has demonstrated that acoustic features extracted from speech signals carry clinically relevant information related to disease severity [–]. However, much of the existing literature on speech-based PD assessment adopts a cross-sectional perspective, treating individual recordings as independent samples. This assumption neglects the longitudinal structure of telemonitoring data, where repeated observations are collected from the same individual over time []. Ignoring temporal dependencies risks conflating inter-patient variability with disease progression and may lead to overly optimistic performance estimates. Moreover, global models that pool all patients implicitly assume a shared disease trajectory, despite well-documented heterogeneity in PD onset, progression rate, and symptom manifestation [,]. From a modeling perspective, PD severity is better viewed as a continuously evolving latent disease state that is imperfectly observed through clinical scores and behavioral signals, such as speech. So, the UPDRS score represents a noisy measurement of this latent state rather than an exact ground truth []. Consequently, approaches that produce deterministic point estimates without accounting for uncertainty fail to reflect the subjective and variable nature of clinical assessment. This limitation is particularly important in telemonitoring contexts, where remotely generated predictions may inform clinical decision-making. An additional limitation of existing approaches is the disconnect between continuous severity estimation and clinically meaningful severity categories. Although UPDRS is reported as a numerical score, clinicians often interpret disease status in terms of ordered severity stages (e.g., mild, moderate, severe) []. However, the telemonitoring dataset used in this study contains only continuous motor UPDRS scores and does not provide clinically validated severity-stage labels. Therefore, a data-driven ordinal categorization is used as an auxiliary modeling strategy to exploit the inherent ordering of disease severity. Models that ignore this ordinal structure may yield numerically accurate predictions that are nonetheless clinically inconsistent. Taken together, these considerations highlight a theoretical gap at the intersection of telemonitoring, disease progression modeling, and clinical interpretability. There is a need for methods that (i) explicitly model longitudinal disease dynamics, (ii) account for patient-specific variability, (iii) quantify uncertainty arising from subjective clinical measurements, and (iv) jointly reconcile continuous severity estimation with ordinal clinical staging.