Fatigue Investigation Using Digital Outcomes

Passive compliance with continuous smartwatch data collection, as a measure of feasibility of long-term fatigue monitoring. Participants wear a smartwatch continuously for two months: a Garmin smartwatch (multiple sclerosis group), an Apple Watch (post-COVID-19 condition group), or either device (healthy controls). Expected wear time is the full duration of each participant's individual enrolment (approximately 60 days), so the denominator is calculated per participant rather than as a fixed period; charging periods are not excluded, so values are not expected to reach 100%. Compliance is calculated per participant as recorded wear time divided by expected wear time, multiplied by 100. Values range from 0% to 100%, with higher values indicating greater compliance. Compliance is also compared by group, disease characteristics and demographic factors, and baseline characteristics associated with lower compliance are described.

Passive compliance with continuous armband data collection, as a measure of feasibility of long-term fatigue monitoring. Participants wear an armband activity monitor (ActiGraph LEAP) continuously for two months. All participants wear the same device, so compliance can be compared between groups independently of device type. Expected wear time is the full duration of each participant's individual enrolment (approximately 60 days), so the denominator is calculated per participant rather than as a fixed period; charging periods are not excluded, so values are not expected to reach 100%. Compliance is calculated per participant as recorded wear time divided by expected wear time, multiplied by 100. Values range from 0% to 100%, with higher values indicating greater compliance. Compliance is also compared by group, disease characteristics and demographic factors, and baseline characteristics associated with lower compliance are described.

Active compliance with self-reported data collection, as a measure of feasibility of long-term fatigue monitoring. Participants complete short questionnaires in the study smartphone application (FIDO app) over two months, comprising cognitive and motor fatigue visual analogue scales prompted four times daily at random times, morning and evening sleep surveys daily, and the Fatigue Severity Scale weekly in the multiple sclerosis and healthy control groups. The number scheduled follows each participant's individual enrolment (approximately 60 days). Compliance is calculated per participant as the number of questionnaires completed divided by the number scheduled, multiplied by 100. Values range from 0% to 100%, with higher values indicating greater compliance. Compliance is reported per questionnaire type, compared by group, disease characteristics and demographic factors, and baseline characteristics associated with lower compliance are described. Response latency is also evaluated.

Active compliance with task-based data collection, as a measure of feasibility of long-term fatigue monitoring. Participants complete two brief performance tasks in the FIDO app approximately every second day over two months: a cognitive fatigability test (cFAST) and a 30-second finger tapping task, together taking approximately seven minutes. The number scheduled follows each participant's individual enrolment (approximately 60 days). Compliance is calculated per participant as the number of sessions completed divided by the number scheduled, multiplied by 100. A session counts as completed if the participant finishes the full task protocol. Values range from 0% to 100%, with higher values indicating greater compliance. Compliance is reported per task, compared by group, disease characteristics and demographic factors, and baseline characteristics associated with lower compliance are described. Response latency is also evaluated.

Participants complete a cognitive fatigability test (cFAST) and a finger tapping test at the study visits and approximately every second day over two months. Performance measures are derived from each test during analysis and are not pre-specified. They are evaluated for their ability to discriminate patients from healthy controls, reported as the area under the receiver operating characteristic curve, from 0.5 (chance) to 1.0 (perfect discrimination), per test. Results are stratified by self-reported fatigue level on the corresponding subscale of the Fatigue Scale for Motor and Cognitive Functions (cognitive subscale for cFAST, motor subscale for tapping; scores 10 to 50, higher indicating greater fatigue). Baseline factors expected to influence this endpoint are age and disease status.

Reliability of smartphone-based performance tests for repeated long-term monitoring. Performance measures are derived during analysis and are not pre-specified. Each test is administered twice in immediate succession, in counterbalanced order, at each of the two study visits, when fatigue is expected to remain stable, and repeated approximately every second day over two months, when fatigue is expected to fluctuate. Agreement across repeated administrations is reported as the intraclass correlation coefficient, from 0 to 1, higher values indicating greater consistency. The immediate-repeat estimate reflects measurement reliability; the two-month estimate reflects the proportion of variance attributable to stable differences between participants. Whether within-participant variation tracks concurrent fatigue is addressed by the corresponding association outcome. Results are reported per test. Baseline factors expected to influence this endpoint are age and disease status.

Agreement between smartphone-based performance tests and established fatigue and functional assessments. Performance measures are derived during analysis and are not pre-specified. A cognitive or motor fatigue visual analogue scale (0 to 10) is completed immediately before every test session. Performance is additionally compared with the fatigue and functional assessments specified in the protocol, including the Fatigue Severity Scale weekly in the multiple sclerosis and healthy control groups, the Fatigue Scale for Motor and Cognitive Functions at the study visits and at four weeks, the Symbol Digit Modalities Test, the Brief Visuospatial Memory Test-Revised at the initial visit, handgrip strength and the Nine-Hole Peg Test. Associations are reported as correlation coefficients from -1 through 0 to +1, per test pairing. Baseline factors expected to influence this endpoint are age and disease status.

Association between digital measurements derived from continuously recorded wearable data and the Fatigue Scale for Motor and Cognitive Functions (FSMC). Candidate measurements are derived during analysis from raw sensor signals recorded by an armband monitor and consumer smartwatch, including photoplethysmography, inertial, barometric and temperature signals; their scales are not pre-specified. Associations are reported as correlation coefficients from -1 through 0 to +1, separately for each group.

Ability of digital measurements derived from continuously recorded armband data to distinguish patients from healthy controls. Candidate measurements are derived during analysis from raw sensor signals recorded by the armband activity monitor (ActiGraph LEAP); their scales are not pre-specified. Only armband-derived measurements are used, because all participants wear the same armband device, whereas the smartwatch differs between groups and would confound comparison between them. Discriminative accuracy for distinguishing patients from healthy controls is reported as the area under the receiver operating characteristic curve, from 0.5 (chance) to 1.0 (perfect discrimination), per derived measurement. Results are stratified by self-reported fatigue level on the Fatigue Scale for Motor and Cognitive Functions (total 20 to 100, higher indicating greater fatigue).

Evaluation of machine learning models developed during the project to estimate fatigue and disease characteristics from digital measurements derived from wearable and smartphone data. Models are newly developed for the study's research objectives; no existing models are validated and none is intended for clinical use. Model families and features are not pre-specified. For classification tasks, performance is reported as the area under the receiver operating characteristic curve, from 0.5 (chance) to 1.0 (perfect discrimination). For regression tasks, performance is reported as prediction error against self-reported fatigue scores, with lower error indicating better performance. Performance is estimated using cross-validation with participant-level splits; for personalised models, splits are made within participants over time. Results are reported separately for each group.

Whether continuously recorded wearable data can detect autonomic dysfunction. At both study visits participants undergo a standardised autonomic test battery based on Ewing's protocol, recorded using a 5-lead ECG Holter and blood pressure monitor, and complete COMPASS-31 (weighted total 0 to 100, higher indicating greater symptom burden). Metrics derived from continuous photoplethysmography and other armband and smartwatch signals, which are not pre-specified, are compared with these measures and reported as measures of association, from -1 through 0 to +1. Discriminative accuracy for identifying participants classified as having autonomic dysfunction on the battery is reported as the area under the receiver operating characteristic curve, from 0.5 (chance) to 1.0 (perfect discrimination). Autonomic comparisons are reported here rather than under the clinical assessment agreement outcome. Results are reported separately for each group, within which the smartwatch is the same device.

Relationship between sleep and self-reported fatigue. Sleep is assessed subjectively through morning and evening surveys completed daily in the FIDO app, and objectively through measures derived from continuous armband and smartwatch recording, including sleep duration, efficiency, timing and estimated sleep stages; these are derived during analysis and not pre-specified, and stages are estimated rather than measured directly. Both are related to fatigue on the cognitive and motor visual analogue scales (0 to 10, higher indicating greater fatigue) and the fatigue questionnaires collected during the study; the interval between sleep and the fatigue reports it is related to is not pre-specified. They are also compared with each other. Relationships are reported as measures of association, from -1 through 0 to +1, and agreement between sleep measures as the mean difference between methods. Results are reported separately for each group.

Relationship between continuously recorded digital measurements and self-reported state fatigue. Participants rate cognitive and motor fatigue on a visual analogue scale (0 to 10, higher indicating greater fatigue) four times daily at random times. Digital measurements derived from armband and smartwatch recording, together with lifestyle data logged by participants in the FIDO app at their own initiative, including diet, activity and symptoms, are related to these ratings; the interval between a measurement and the ratings it is related to is not pre-specified, so that measurements preceding a fatigue rating can be examined as possible precursors. Sleep measures and smartphone-based test performance are reported under their own outcomes. Relationships are reported as measures of association, from -1 through 0 to +1, separately for each group.

Assessment of post-exertional malaise, defined as worsening of symptoms following exertion, in participants with post-COVID-19 condition. Participants log symptoms in the FIDO app at their own initiative, rating severity as low, moderate or high. Measures of exertion and physiological response are derived during analysis from continuous armband and smartwatch recording, including activity, photoplethysmography-derived metrics, blood pressure and skin temperature, and from self-logged activity. These are related to subsequently reported symptom occurrence and severity and to fatigue visual analogue scale ratings (0 to 10, higher indicating greater fatigue), and are considered relative to each participant's usual level; neither the threshold defining exertion nor the interval before subsequent reports is pre-specified. Time since infection, medications and sleep quality are considered as modifying factors. Relationships are reported as measures of association, from -1 through 0 to +1.

Evaluation of machine learning models developed during the project to predict short-term fluctuations in self-reported state fatigue from digital measurements and lifestyle reporting. Models may use measurements recorded before the fatigue rating being predicted, to assess whether objective signals precede fatigue episodes. Models are newly developed for the study's research objectives; no existing models are validated and none is intended for clinical use. Model families and features are not pre-specified. For classification of high versus low fatigue states, performance is reported as the area under the receiver operating characteristic curve, from 0.5 (chance) to 1.0 (perfect discrimination). For continuous prediction of fatigue ratings (0 to 10), performance is reported as prediction error, with lower error indicating better performance. Cross-validation uses participant-level splits. Results are reported separately for each group.

Association between digital measurements derived from continuously recorded wearable data and the Fatigue Severity Scale (FSS). Candidate measurements are derived during analysis from raw sensor signals recorded by an armband monitor and a consumer smartwatch, including photoplethysmography, inertial, barometric and temperature signals; their scales are not pre-specified. The FSS is administered only to participants in the multiple sclerosis and healthy control groups; associations are reported for these groups as correlation coefficients from -1 through 0 to +1.

Post hoc analysis of how the amount, resolution and completeness of recorded data affect the digital biomarker outcomes for fatigue and disease characteristics. The recorded dataset is resampled at reduced temporal resolutions, with increasing proportions of data removed, and truncated to shorter recording durations, simulating lower-specification devices, less frequent synchronisation and shorter wear periods. Digital measurements derived from these reduced datasets are compared with those derived from the full dataset, and agreement is reported as correlation coefficients from -1 through 0 to +1. For truncation, the outcome is additionally the number of days after which a measurement no longer changes by more than a predefined margin. Machine learning models trained on the reduced datasets are compared with those trained on the full dataset using the same metrics. Results are reported separately for each group, and inform the design of data collection in future studies.

Assessment of whether menstrual cycle phase is associated with fatigue in participants who menstruate. The dates of the most recent menstrual period are recorded at the initial visit and cycle dates covering the recording period by questionnaire at the final visit; phase is estimated from these as follicular or luteal, with continuous wearable measurements used to corroborate self-reported phase. Hormonal contraceptive use is recorded at baseline and accounted for, as it may suppress or alter cyclical variation; perimenopausal and postmenopausal participants are considered separately; and cycle regularity may be affected by the conditions studied, which is taken into account when interpreting phase estimates. Phase is related to fatigue on the visual analogue scale (0 to 10) and the Fatigue Severity Scale (score 1 to 7), higher values indicating greater fatigue, and to digital measurements. Associations are reported as measures of association, from -1 through 0 to +1.

Characterisation of gut microbiome composition and its variation between clinical groups. Participants collect stool samples at four and eight weeks after the initial visit, which are analysed for microbial composition, summarised as within-sample (alpha) diversity, between-sample (beta) diversity and relative abundance of bacterial taxa. Differences in these measures are examined between participants with multiple sclerosis, participants with post-COVID-19 condition and healthy controls. Participant characteristics that may alter composition are recorded and reported, including medication use such as antibiotics, dietary and lifestyle changes recorded through the diary function of the FIDO app, body mass index, time since disease onset or acute SARS-CoV-2 infection, gastrointestinal symptoms and comorbidities. Values are averaged across the two sampling points to give a more stable estimate; change between them is also examined.

Perceived usability and quality of the FIDO app. Before the final visit participants complete the User Experience Questionnaire, yielding six scale scores from -3 to +3 where higher values indicate a more positive experience, and the Mobile Application Rating Scale, yielding four objective quality subscales and a subjective quality subscale, scored 1 to 5 where higher values indicate better quality. The subjective quality subscale covers willingness to recommend the app, expected frequency of future use and overall rating. Associations between usability scores and compliance with scheduled questionnaires and performance tasks are reported as correlation coefficients from -1 through 0 to +1. Results are reported separately for each group.

Relationship between gut microbiome composition, fatigue and autonomic function. Stool samples collected at four and eight weeks are analysed for microbial composition, summarised as diversity and relative abundance of bacterial taxa, averaged across the two sampling points. These measures are related to fatigue on the Fatigue Scale for Motor and Cognitive Functions (total 20 to 100) and the Fatigue Severity Scale in the multiple sclerosis and healthy control groups (1 to 7), higher values indicating greater fatigue; to autonomic function assessed by COMPASS-31 and a standardised autonomic test battery; to digital measurements derived from wearable data; and to diet and symptom entries logged in the FIDO app. Participant characteristics that may influence composition are reported alongside the associations rather than adjusted for. Associations are reported as measures of association, from -1 through 0 to +1, separately for each group.

Assessment of whether participation in the study, which involves no intervention, was associated with a change in participants' everyday behaviour. Lifestyle metrics recorded by participants' own devices before enrolment, including step count and other activity and sleep measures, are extracted retrospectively from Apple Health for the two months preceding the initial visit and compared with the corresponding metrics recorded during the two-month data collection period. Change is reported as the difference between the two periods for each metric, with positive values indicating an increase during the study. This analysis is possible only for participants who use an Apple device and who consent to sharing historical data, and results are therefore reported for that subgroup and separately for each group.

Participant experience of using the FIDO app and wearable devices over two months. At the final visit participants complete a semi-structured interview in German or English, which is audio recorded and transcribed. Transcripts are analysed thematically, covering intention to continue using the app and devices beyond the study, moments of disengagement, barriers to sustained use, sources of motivation, perceived fit of the study design to individual needs, and experience of the wearable devices. Participants also report whether they owned wearable devices before the study and which device placement they preferred, upper arm or wrist; these are reported as counts and proportions. Themes are reported with frequency across participants and illustrative anonymised quotations, examined for differences between the multiple sclerosis, post-COVID-19 condition and healthy control groups, and compared with usability scores and compliance data.

Association between digital measurements derived from continuously recorded wearable data and the Munich Berlin Symptom Questionnaire (MBSQ). Candidate measurements are derived during analysis from raw sensor signals recorded by an armband monitor and consumer smartwatch, including photoplethysmography, inertial, barometric and temperature signals; their scales are not pre-specified. The MBSQ is administered only to participants in the Post-COVID-19 group; associations are reported for this group as correlation coefficients from -1 through 0 to +1.

Association between digital measurements derived from continuously recorded wearable data and the Bell Disability Scale. Candidate measurements are derived during analysis from raw sensor signals recorded by an armband monitor and consumer smartwatch, including photoplethysmography, inertial, barometric and temperature signals; their scales are not pre-specified. The Bell Disability Scale is administered only to participants in the Post-COVID-19 group; associations are reported for this group as correlation coefficients from -1 through 0 to +1.

Association between digital measurements derived from continuously recorded wearable and smartphone data and the Symbol Digit Modalities Test (SDMT). Candidate measurements are derived during analysis from raw sensor signals recorded by an armband monitor and consumer smartwatch, including photoplethysmography, inertial, barometric and temperature signals, and from smartphone-based measurements collected through the study application. Their scales are not pre-specified. Associations are reported as correlation coefficients from -1 through 0 to +1, separately for each group.

Association between digital measurements derived from continuously recorded wearable data and from smartphone-based assessments and the Brief Visuospatial Memory Test-Revised (BVMT-R). Candidate measurements are derived during analysis from raw sensor signals recorded by an armband monitor and consumer smartwatch, including photoplethysmography, inertial, barometric and temperature signals, and from smartphone-based active tasks and questionnaires administered through the study application; their scales are not pre-specified. Associations are reported as correlation coefficients from -1 through 0 to +1, separately for each group.

Characterisation of autonomic function and its variation between clinical groups. At both study visits participants undergo a standardised autonomic test battery based on Ewing's protocol, comprising heart rate response to paced deep breathing at 0.1 Hz, the 30:15 ratio on active standing, blood pressure response to postural change, and blood pressure response to sustained handgrip at 30% of maximal voluntary contraction, recorded using a 5-lead ECG Holter and blood pressure monitor. Autonomic symptoms are assessed by COMPASS-31 (weighted total 0 to 100, higher indicating greater symptom burden). Values from each component of the battery are compared between participants with multiple sclerosis, participants with post-COVID-19 condition and healthy controls.

Whether blood pressure estimated from photoplethysmography by the armband activity monitor (ActiGraph LEAP), worn on the upper arm, agrees with cuff-measured blood pressure. Comparison is made only at the study visits, during the autonomic test battery, when cuff blood pressure is recorded in the supine position, at one, three and five minutes of standing, and before, during and after a five-minute sustained handgrip contraction at 30% of maximal voluntary contraction. Estimated and cuff-measured systolic and diastolic blood pressure are compared at each of these timepoints, in millimetres of mercury. Agreement is assessed using Bland-Altman analysis and reported as the mean difference between methods with 95% limits of agreement. Agreement during postural change and sustained contraction, when blood pressure is expected to vary, is examined separately.

Responsiveness to scheduled prompts, as a measure of the effort required to sustain active data collection rather than whether it is completed at all. The FIDO app records the time at which each prompt is sent and the time at which the corresponding questionnaire or task is submitted; latency is the interval between these timestamps, in minutes, calculated per completed prompt, with longer intervals indicating slower responding. Prompts that are never completed are not assigned a latency and are captured by the completion rate outcomes. Latency is summarised per participant as the median and as within-participant variability over the two-month period, reported separately for each prompt type (fatigue visual analogue scale, sleep surveys, Fatigue Severity Scale, cognitive fatigability test, finger tapping task), and compared by group, disease characteristics and demographic factors.