Movement as a Neurological Biomarker
ABSURD Health Institute · Volume Two · Research Essay 54
Research series: 08 · Signals, measurement and responsible medical AI
Abstract
Gait, tremor and fine motor movements are observable manifestations of nervous-system function. Video and wearable sensors may capture frequency, amplitude, speed and asymmetry more consistently than informal descriptions. This paper develops movement tracking as a longitudinal biomarker while warning that a measured tremor signal is not a diagnosis of Parkinson disease or another neurological disorder.
The motor phenotype as repeated measurement
Gait and tremor recordings can reveal neurological change invisible to a static examination, provided the measurements are specific enough to be useful.
Movement disorders require phenomenology: rest versus action tremor, rhythmicity, posture, task dependence, accompanying slowness, stiffness or ataxia. Video can help document some features repeatedly, but camera angle, frame rate, lighting, background and voluntary effort influence analysis. Rigidity cannot be directly assessed in a standard video, and clinical bradykinesia involves decrement and neurological examination beyond mere movement speed.
A research-worthy device would standardize tasks, document the conditions of capture and compare measurements against specialist assessments. It could track within-person change during medication adjustment or rehabilitation, provided confounding from fatigue, anxiety and time of day is considered. Capturing a movement abnormality is a useful measurement; assigning its cause requires clinical history, examination and often additional testing.
What gait and tremor features signify
Tremor frequency can sometimes be quantified more consistently from accelerometer or video signals than from memory, but frequency alone is rarely diagnostic. The same person may display different movement patterns during rest, posture, finger tapping and purposeful action. Camera-based tools may capture amplitude and velocity but cannot reliably infer every element of a neurological examination. Stress, caffeine, sleep deprivation and medication timing change movement. A standardized capture workflow is therefore inseparable from the meaning of the output. Without it, apparent disease progression might simply reflect different filming conditions.
Movement captures the interaction of the nervous system, musculoskeletal structures, energy availability and motivation. Changes in walking speed, step variability, tremor or balance can suggest deterioration but are not uniquely diagnostic. A slow gait might result from pain, fatigue, neuropathy, Parkinsonian features or simply carrying a heavy object. The most informative signals may be change from a well-defined personal baseline under comparable conditions rather than deviation from a single age-adjusted norm.
Wearable sensors and video analysis can measure different dimensions. A wrist accelerometer may detect rhythmic motion but not reliably distinguish essential tremor from Parkinson's disease without additional clinical context. Smartphone video can document amplitude, task dependence and progression, but camera position and frame rate shape the result. Labels supplied by an experienced clinician are needed to validate phenotype recognition.
Step variability, movement amplitude, freezing, asymmetry and tremor frequency reflect interacting motor circuits and musculoskeletal constraints. A consumer video may capture clinically relevant changes but cannot reliably classify all movement disorders without history and expert examination. The strong thesis is that standardized longitudinal video can quantify decline or response more sensitively than occasional subjective description.
Separating progression from recording artifacts
A prospective study should ask whether sensor-derived trajectories predict deterioration before conventional examination and whether an alert leads to useful intervention. It should distinguish a device's ability to classify prerecorded examples from its performance on unselected patients with comorbidities. Patients with new sudden weakness, trouble walking or focal neurological deficits should receive urgent assessment rather than wait for repeated device observations to confirm a statistical anomaly.
Different movements demand different features. Tremor is characterized by frequency, amplitude, regularity, activation condition and distribution; gait may require stride variability, velocity and turning performance. A general “movement wellness score” could obscure the features that neurologists actually use. Video analysis must distinguish rest, posture and action, while wearable analysis needs enough sensor resolution and an understanding of where the device is worn.
The medication problem. Changes in antipsychotic, stimulant or antiparkinsonian medications can shift movement patterns, as can sleep loss and illness. If a model predicts diagnosis partly by medication-correlated signals, it may fail to identify people before treatment. Validation cohorts should include medication-naive and treated patients where feasible and report performance separately.
Clinical use and limitations. Remote monitoring may be most useful for tracking an already characterized neurological disorder or identifying significant change from baseline. It cannot independently confirm that a tremor originates in a particular brain circuit or that a long-standing diagnosis is mistaken. A decision support tool that produces clear measurements for a clinician could still be highly valuable without claiming autonomous diagnosis.
Compare repeated video-derived features with blinded neurologic ratings and clinical outcomes, testing within-person reproducibility and generalization across cameras. Include control conditions such as pain, fatigue and medication state. A classifier that looks accurate only when evaluating the same patient’s training videos is not clinically validated.
The critical validation question is whether the digital measurement corresponds to a clinically meaningful endpoint such as specialist-rated severity or longitudinal functional change. If a model accurately classifies tremor frequency but fails to distinguish essential tremor from parkinsonian tremor, it should be described as a measurement aid rather than a diagnostic system. Clear labeling protects patients and improves clinician trust.
The important signal may be variation rather than the average movement
Two people may walk at the same average pace while one has increasing step-to-step variability, asymmetry or freezing episodes. Likewise the mean amplitude of a tremor may conceal a change in frequency or its response to posture and medication. These details can be quantified with standardized video or wearable sensors, but the measurements depend on camera angle, frame rate, footwear, fatigue and the testing task. A model trained on carefully supervised recordings may perform poorly when a patient films themselves in an ordinary room. A strong study should therefore measure repeatability under varied conditions, distinguish changes caused by pain from neurological progression and compare algorithmic outputs with blinded specialist ratings. Its most valuable prediction would be clinically meaningful change before routine appointments detect it. Quantification is not the same as diagnosis: different diseases may share movement patterns, and the final etiological assessment still requires history and examination.
Video complements the examination
Movement classification must distinguish the physical pattern from the illness that may produce it. Similar tremor frequencies or walking changes can arise from neurological disease, medication, fatigue, pain and normal variation. A machine-learning system might recognize a waveform reliably while remaining unable to identify its cause. Studies should label the movement task, medication state, sensor location and recording quality before making comparisons. Longitudinal patterns may be especially valuable because the same person's baseline reduces some between-person variation, though changes in device or environment can create artifacts. A validated measurement of tremor severity could help a clinician follow a known diagnosis; that does not authorize the model to overturn the diagnosis. The research should first show accurate measurement and then separately test whether that measurement improves clinical decisions. New sudden neurological deficits require emergency care irrespective of any previously reassuring sensor score.
The design should also preserve the difference between within-person monitoring and between-person diagnosis. A model may track one person’s established tremor consistently yet fail to differentiate two diseases in a new patient. This distinction should be reflected in intended-use claims, validation cohorts and the wording of alerts. A clinically useful product does not need to overstate its diagnostic reach.
Algorithm performance may look strong in a curated dataset of severe disease and fail among patients with subtle presentations, essential tremor or medication-induced symptoms. Labels inherited from inconsistent clinical diagnoses can further distort training. Even perfect recognition of a tremor frequency does not prove an underlying neurodegenerative pathology.
The strongest application is objective, standardized description and follow-up, not unsupervised diagnosis. Signals could frame movement tracking as a way to help patients bring better longitudinal information to clinicians, with privacy, consent and clear escalation guidance for new neurological deficits.
Sources and research leads
Editorial status: Third-pass conceptual research manuscript. The argument has been individually strengthened and shared boilerplate removed. Original reference leads were retained; a complete independent source-by-source scholarly verification has not been performed. No new clinical trial data are presented.