Introduction
Parkinson disease is a progressive neurodegenerative disorder characterized by increasing motor and nonmotor symptoms, declining quality of life, and greater health care use (1–Reference Armstrong and Okun4). As the disease advances, symptoms may become insufficiently controlled with conventional pharmacological treatment, and selected patients may benefit from device-aided therapies such as deep brain stimulation (DBS) or infusion therapies (Reference Antonini, Pahwa and Odin5). A key clinical challenge is that progression to advanced Parkinson disease (APD) may not be recognized promptly during routine follow-up because symptoms are heterogeneous, fluctuate over time, and may not be fully captured during periodic clinical visits (Reference Malaty, Martinez-Martin and Chaudhuri6). Delayed recognition can postpone treatment optimization and referral for advanced therapy assessment.
Remote monitoring systems may help address this gap by collecting longitudinal information on symptoms and functioning outside the clinic (Reference Godinho, Domingos and Cunha7;Reference Burq, Rainaldi and Ho8). Wearable and nonwearable devices, electronic symptom diaries, and self-recorded motor assessments can provide repeated data on mobility, sleep, activity, and symptom fluctuations. Combined with machine learning (ML), these data may support earlier recognition of disease progression and generate alerts for clinician review (Reference Landers, Saria and Espay9–Reference Islam, Rahman and Abdelkader11). However, the clinical utility and economic value of ML-enabled remote monitoring systems (ML-RMS) remain uncertain. Diagnostic accuracy alone may be insufficient if alerts do not change clinical decisions, treatment timing, patient outcomes, or long-term care needs.
Early health technology assessment (HTA) is well suited to evaluate such uncertainty during technology development (Reference Grutters, Bouttell and Abrishami12). Rather than providing a definitive reimbursement conclusion, early HTA can estimate potential value, identify conditions under which a technology may become cost-effective, and prioritize evidence gaps for further research and validation (Reference Buisman, Rutten-van Mölken, Postmus, Luime, Uyl-de Groot and Redekop13–Reference Elsisi, Canestaro, Steuten and Hansen15).
This study conducted an early HTA of an ML-RMS designed to support earlier APD recognition. From the Finnish health care perspective, we estimated its potential cost-effectiveness compared with standard care and identified key value drivers, development targets, and evidence gaps for future validation and implementation.
Methods
A model-based early HTA estimated the potential cost-effectiveness of ML-RMS to support earlier recognition and confirmation of APD compared with standard care. Because the technology was still under development, the analysis was designed to support development, validation, and evidence-generation decisions. The evaluation followed established health economic evaluation guidance (Reference Drummond and Jefferson16) and the CHEERS-AI reporting framework (Supplementary file 1) (Reference Elvidge, Hawksworth and Avşar17).
The decision problem reflected the technology’s early development stage, intended clinical use, and the model’s role as a development-support tool (Table 1). The analysis examined whether ML-RMS could generate value by enabling earlier APD recognition and, in eligible patients, timely DBS assessment and treatment.
Decision problem for the early health technology assessment
Table 1. Long description
The table consists of two columns: Element and Description.
* Technology: M L R M S for detecting progression to advanced Parkinson disease.
* Development stage: Early proof-of-concept; Technology Readiness Level 3.
* Intended use: To support clinicians by identifying patients with possible progression to advanced Parkinson disease.
* Target population: Persons with Parkinson disease at increased risk of progression to advanced disease.
* Intervention: M L R M S-supported care, in which remote monitoring data are analyzed to generate alerts for possible disease progression.
* Comparator: Standard care with routine in-person clinical follow-up.
* Perspective: Finnish health care perspective.
* Time horizon: 16 years.
* Model type: Markov state-transition model.
* Main outcomes: Incremental costs, Q A L Y s, I C E R, and i N M B.
* Main value mechanism: Earlier recognition of advanced Parkinson disease enabling timely D B S in eligible patients. Broader treatment optimization was explored through scenario analyses.
* Decision informed: Whether further development and validation are justified, and which parameters require additional evidence generation.
* Key uncertainties: Diagnostic accuracy, clinician response to alerts, D B S eligibility and timing, quality-of-life effects, nursing home admission, implementation costs, adherence, and monitoring burden.
DBS, deep brain stimulation; ICER, incremental cost-effectiveness ratio; iNMB, incremental net monetary benefit; ML-RMS, machine learning-enabled remote monitoring system; QALY, quality-adjusted life year.
Technology and intended use: ML-RMS
The ML-RMS is being developed and piloted by Helsinki University Hospital (HUS) with clinical and technical partners within the EU-funded AICCELERATE project (18). It targets persons with Parkinson disease at increased risk of progression to APD and is intended to support, rather than replace, clinical judgment by generating alerts for possible disease progression. Diagnostic confirmation and treatment decisions remain the responsibility of health care professionals.
The intended ML-RMS is designed to combine data from wearable sensors, self-recorded motor assessments, electronic symptom diaries, and electronic health records to identify patterns suggestive of APD progression. It does not directly diagnose APD or determine eligibility for advanced therapies. Instead, alerts prompt clinician review and, when appropriate, specialist referral. Patients with confirmed APD are then assessed for DBS or other advanced therapies according to usual clinical criteria.
At the time of evaluation, the ML-based predictive model was at an early proof-of-concept stage (Technology Readiness Level 3) (Reference Mankins19). The preliminary model had been developed and internally evaluated using electronic health record data, while incorporation and prospective validation of additional remote-monitoring data streams were ongoing. Final algorithm architecture, training data, and definitive performance metrics were not yet available. The remote-monitoring component was also undergoing usability testing and prospective data collection to support further development.
Model-based economic evaluation
A preliminary health economic analysis plan (Supplementary file 2) guided the model-based evaluation. The analysis was informed by iterative input from two neurologists at HUS involved in the AICCELERATE project: L.M., a specialist neurologist with expertise in digital health and remote monitoring in Parkinson disease, and M.K., a specialist neurologist with expertise in movement disorders, APD, and DBS. They were selected based on their relevant clinical expertise and knowledge of the intended care pathway. Their input informed assumptions on clinician response to ML-RMS alerts, specialist referral, APD confirmation, and eligibility for advanced therapies. No patients, members of the public, or other multidisciplinary stakeholders were involved. No conflicts of interest were declared, and uncertain expert-informed assumptions were explored through sensitivity and scenario analyses.
The model simulated a hypothetical cohort of 1,000 persons with Parkinson disease using a Markov state-transition model with four mutually exclusive health states: (i) non-APD at risk of APD; (ii) diagnosed APD living at home; (iii) APD residing in a nursing home; and (iv) death. Transitions were modeled in 6-month cycles to reflect routine clinical follow-up and capture clinical and cost-related outcomes over the disease course. Figure 1 shows the model structure. Table 2 summarizes key disease progression timelines and assumptions, and Figure 2 presents the clinical pathway linking APD recognition, diagnostic confirmation, DBS eligibility, and downstream outcomes. The model was developed and analyzed in Microsoft Excel for Microsoft 365, version 2607 (Microsoft Corporation, Redmond, WA, USA).
Markov model structure. Markov state-transition model simulating disease progression in Parkinson disease under ML-RMS-supported care and standard care. Abbreviations: ML-RMS, machine learning-enabled remote monitoring system; PD, Parkinson disease.
Disease progression timelines and key model assumptions for ML-RMS and standard care
Table 2. Long description
The table consists of three columns: Parameter, Value, and Source.
* Survival difference between M L - R M S and standard care: Value is No difference; Source is Crispo et al. 20, Mahlknecht et al. 21.
* Prevalence of A P D at 6.6 years after P D onset: Value is 36 percent; Source is Martínez-Castrillo et al. 22.
* Estimated prevalence of A P D at 8 years after P D diagnosis: Value is 55.2 percent; Source is Weibull extrapolation.
* Sensitivity/specificity of routine clinical assessment: Value is 70 percent / 60 percent; Source is Expert-informed assumption, informed by A P D screening literature 6, 23.
* Sensitivity/specificity of M L - R M S: Value is 88 percent / 62 percent; Source is Early retrospective internal validation evidence.
* Average D B S initiation time after P D diagnosis under standard care: Value is 10.83 years; Source is Model calibration to literature-based target of approximately 11 years after P D diagnosis, de Souza et al. 24.
* Probability of receiving D B S after A P D diagnosis: Value is 10 percent; Source is Expert opinion.
* Time to nursing home admission after P D diagnosis: Value is 16 years; Source is Crispo et al. 20, Vossius et al. 25.
Abbreviations defined at the bottom: A P D, advanced Parkinson disease; D B S, deep brain stimulation; M L - R M S, machine learning-enabled remote monitoring system.
APD, advanced Parkinson disease; DBS, deep brain stimulation; ML-RMS, machine learning-enabled remote monitoring system.
Modeled clinical pathways for standard care and ML-RMS-supported care in advanced Parkinson disease. The figure illustrates how ML-RMS was represented as a clinical decision-support tool rather than a diagnostic replacement. In each 6-month cycle, diagnostic performance was applied to patients who were alive and not yet clinically recognized as having APD. In the ML-RMS-supported pathway, a positive alert prompted clinician review and specialist assessment, followed by confirmed APD diagnosis after a diagnostic-to-confirmation delay. Patients with confirmed APD were subsequently assessed for DBS eligibility; in the base case, 10 percent were assumed to receive DBS. Patients not yet recognized as having APD remained eligible for detection in subsequent cycles. APD, advanced Parkinson disease; DBS, deep brain stimulation; ML-RMS, machine learning-enabled remote monitoring system.
Figure 2. Long description
Panel A, Standard care pathway. The flow begins with People with Parkinson disease 8 years after diagnosis. This leads to a Routine 6-month clinical assessment. A decision diamond asks if A P D is clinically recognized. If No, the patient remains eligible for detection in later cycles and returns to the routine assessment. If Yes, the path leads to Specialist assessment and confirmed A P D diagnosis after diagnostic delay, followed by D B S eligibility assessment. A second decision diamond asks if the patient is Eligible for D B S. If Yes, they receive D B S. If No, they receive Best medical therapy. Both paths lead to Diagnosed A P D living at home, then Nursing home admission, and finally Death.
Panel B, M L - R M S-supported pathway. The flow is similar but starts with M L - R M S monitoring plus routine care. The first decision diamond asks if there is an M L - R M S alert for possible A P D. If No, they remain eligible for later cycles. If Yes, it leads to Clinician review before proceeding to Specialist assessment and the same subsequent steps as Panel A.
Model notes at the bottom are numbered 1 through 4. Note 1 states diagnostic performance was applied in each 6-month cycle. Note 2 explains positive M L - R M S assessments prompted clinician review rather than immediate diagnosis. Note 3 states 10 percent of patients with confirmed A P D received D B S. Note 4 mentions scenario analysis explored quality-of-life gains and delayed nursing home admission.
The simulated cohort consisted of persons with Parkinson disease at increased risk of progression to APD, with a mean age of 68 years (Reference Armstrong and Okun4). Patients were assumed to enter the model 8 years after Parkinson disease diagnosis, representing a population in whom monitoring for progression to APD would be clinically relevant. Progression to APD was estimated using prevalence data from Martínez-Castrillo et al. (Reference Martínez-Castrillo, Martínez-Martín and Burgos22), who reported that 36 percent of persons with Parkinson disease had APD 6.6 years after disease onset. Because available progression evidence was reported by disease duration from onset, while the model was structured by time since diagnosis, disease onset and diagnosis were assumed to occur at the same time.
Progression probabilities over time were extrapolated using a Weibull distribution to represent a gradual increase in the probability of APD with increasing disease duration. In the model, all surviving patients were assumed to have progressed to APD by 16 years after Parkinson disease diagnosis (Reference Crispo, Lam and Le20), reflecting the advanced-risk cohort and the advanced disease stage captured over the analytical time horizon (Supplementary file 3).
In each cycle, the model distinguished between underlying true APD status and clinically recognized APD status. Incident APD cases were estimated from the extrapolated APD prevalence curve among surviving patients. Patients with undiagnosed APD who were not flagged remained eligible for detection in later cycles, unless they died.
In the ML-RMS arm, patients began remote monitoring 8 years after Parkinson disease diagnosis. ML-RMS was assumed to have sensitivity and specificity of 88 percent and 62 percent, respectively. These estimates were derived from preliminary internal validation of the EHR-based proof-of-concept model and should therefore be interpreted as early development stage inputs rather than performance estimates for the final multimodal ML-RMS or routine clinical use. Because direct estimates of routine clinical recognition of APD are unavailable, standard care diagnostic performance was modeled using conservative expert-informed assumptions, informed by published studies, with sensitivity and specificity set at 70 percent and 60 percent, respectively (Reference Malaty, Martinez-Martin and Chaudhuri6;Reference Martinez-Martin, Kulisevsky, Mir, Tolosa, García-Delgado and Luquin23).
Diagnostic performance was applied in each cycle to patients who were alive and not yet clinically recognized as having APD. Among patients with undiagnosed APD, sensitivity determined the proportion generating a potential true-positive APD flag. Among patients without APD, one minus specificity determined the proportion generating a false-positive assessment.
The model separated potential APD flagging from confirmed APD diagnosis. Potential true-positive APD flags initiated clinical review and specialist assessment rather than immediate confirmed diagnosis. Patients with a potential true-positive APD flag left the undiagnosed APD pool and entered a pending confirmation pathway. Confirmed APD diagnosis occurred after a distributed diagnostic-to-confirmation delay modeled over 0 to 7 cycles, with weights of 4 percent, 8 percent, 14 percent, 21 percent, 22 percent, 16 percent, 10 percent, and 5 percent, respectively. The mean delay was 3.62 cycles, corresponding to 1.81 years. Survival between flagging and confirmation was accounted for.
This delay distribution was calibrated so that standard care reproduced an average DBS initiation time close to the literature-based target of 11 years after Parkinson disease diagnosis (Reference de Souza, Akram, Low, Green, Ashkan and Schapira24). In the final model implementation, the average DBS initiation time under standard care was 10.83 years. The same diagnostic-to-confirmation delay distribution was applied in the base case ML-RMS arm, so that the base case comparison isolated the effect of improved diagnostic accuracy rather than assuming a faster clinical pathway.
The primary outcomes were total costs, life years, quality-adjusted life years (QALYs), incremental costs and QALYs, incremental cost-effectiveness ratio (ICER), and incremental net monetary benefit (iNMB). ML-RMS was assumed to improve outcomes through earlier APD recognition and confirmation and, among eligible patients, timely DBS initiation, potentially enabling earlier access to its quality-of-life benefits (Reference Deuschl, Schüpbach and Knudsen26).
Based on clinical expert input, 10 percent of patients with confirmed APD were assumed to be eligible for and subsequently receive DBS. This proportion was an expert-informed model assumption. DBS eligibility was assumed to follow usual Finnish clinical practice, consistent with the national evidence-based guideline for Parkinson disease (27). New DBS initiations in each cycle were calculated as a fixed proportion of newly confirmed true-positive APD diagnoses. Patients receiving DBS entered cycle-specific DBS cohorts according to the time of initiation, allowing DBS-related costs and QALYs to accrue over subsequent cycles while accounting for survival. Patients not receiving DBS continued best medical therapy. Other device-aided therapies, such as levodopa infusion therapy, were not modeled separately because of limited early-stage evidence and to preserve model simplicity.
Patients with APD living at home were assumed eventually to transition to nursing home care 16 years after Parkinson disease diagnosis, reflecting published estimates of time from diagnosis to nursing home admission (Reference Crispo, Lam and Le20;Reference Vossius, Nilsen and Larsen25). Nursing home timing was assumed to be the same in standard care and ML-RMS arms in the base case. State transitions were permanent, and no survival difference between ML-RMS and standard care was assumed (Reference Crispo, Lam and Le20;Reference Mahlknecht, Foltynie, Limousin and Poewe21).
Costs and quality-of-life inputs
Costs and health outcomes were accumulated in 6-month cycles until death or for a maximum model duration of 16 years. Health outcomes were expressed as life years and QALYs. Costs and health effects were discounted at an annual rate of 3 percent, in line with Finnish health economic evaluation guidelines. Because Finland does not have an official cost-effectiveness threshold, a willingness-to-pay threshold of €35,000 per QALY gained was used, based on previous Finnish cost-effectiveness analyses (Reference Hallinen, Kivelä, Soini, Harjola and Pesonen28).
Incremental net monetary benefit was calculated as λ × incremental QALYs minus incremental costs, where λ represents the willingness-to-pay threshold. A half-cycle correction was applied (Reference Drummond, Sculpher, Claxton, Stoddart and Torrance29). All costs were expressed in euros and indexed to 2023 values (30;31).
Because the exact costs of ML-RMS were not established, the base case estimate was derived from a published economic evaluation of an existing remote monitoring system (Reference Cox, Wade and Hodgson32). DBS surgery costs were obtained from Dams et al. (Reference Dams, Balzer-Geldsetzer and Siebert33) and adjusted to the Finnish setting. Resource use before and after DBS was informed by the PD SURG trial, which reported no major differences in overall health care resource use between DBS and best medical therapy except for nurse visits, speech and language therapy sessions, and hospital outpatient visits (Reference McIntosh, Gray and Daniels34). Detailed cost inputs, resource-use assumptions, and utility values are provided in Supplementary file 4.
Utility values were derived from published literature. Baseline utility at 8 years after Parkinson disease diagnosis was derived from Zahra et al. (Reference Zahra, Durand-Zaleski, Górecki, Walleser Autiero, Barnett and Schüpbach35). Under standard care, DBS was assumed to be initiated approximately 11 years after Parkinson disease diagnosis, at which point a 22 percent proportional increase in utility was applied relative to the pre-DBS value (Reference Lachenmayer, Mürset and Antih36). Because quality-of-life gains may differ according to the timing of DBS initiation (Reference Schuepbach, Rau and Knudsen37), proportional utility gains of 26 percent, 30 percent, and 34 percent were applied when DBS was initiated at 10, 9, and 8 years after diagnosis, respectively. Because estimates of time-dependent DBS utility gains were unavailable, these larger proportional gains were modeled as exploratory assumptions. These assumptions allowed DBS initiated at earlier modeled time points to generate a larger initial utility gain while maintaining the same post-DBS deterioration slope as in the standard care pathway.
Over time, DBS benefits were assumed to attenuate because of aging and progression of nonmotor symptoms, with utility converging to that of patients managed without DBS by 16 years after diagnosis. Utility values for patients residing in nursing homes were assumed to be identical across strategies. In the base case analysis, ML-RMS itself was not assumed to have a direct effect on quality of life independent of DBS timing.
Sensitivity analysis
Scenario analyses assessed key structural and parameter uncertainties, including diagnostic performance, intervention and DBS costs, implementation timing, clinical response to alerts, long-term care needs, and direct quality-of-life effects. Scenarios increased ML-RMS sensitivity and specificity to 95 percent, reduced ML-RMS or DBS procedure costs by 50 percent, and delayed ML-RMS initiation from 8 to 9 years after Parkinson disease diagnosis.
Because the base case applied the same diagnostic-to-confirmation delay in both arms, a faster-pathway scenario reduced the average time from APD flagging to confirmed diagnosis and DBS assessment by 12 months. Additional scenarios examined a 6-month delay in nursing home admission for 1 percent, 5 percent, or 10 percent of patients in the ML-RMS arm, and direct proportional utility gains of 1 percent, 5 percent, or 10 percent during active monitoring, independent of DBS timing.
The direct utility-gain scenarios represented potential benefits beyond APD recognition, including improved self-management, symptom control, reassurance, or communication with clinicians. Utility gains applied only during active ML-RMS use, which ended after APD diagnosis or nursing home admission, and were considered exploratory.
A PSA quantified joint parameter uncertainty (Reference Drummond, Sculpher, Claxton, Stoddart and Torrance29). Key clinical, cost, and utility parameters were varied simultaneously across 1,000 Monte Carlo iterations. Uncertainty in incremental costs, QALYs, ICER, and iNMB was summarized using a cost-effectiveness plane and cost-effectiveness acceptability curve (Supplementary file 5).
A threshold-style analysis also examined how alternative combinations of ML-RMS sensitivity and specificity affected incremental costs and QALYs and whether improved diagnostic performance alone could make ML-RMS favorable at the selected willingness-to-pay threshold.
Results
In the base case analysis, ML-RMS generated a small health gain of 0.0044 QALYs compared with standard care, while increasing total costs by €761 per patient. The resulting ICER was €172,023 per QALY gained. At a willingness-to-pay threshold of €35,000 per QALY, the iNMB was −€607, indicating that ML-RMS was not cost-effective under base case assumptions. This unfavorable result was primarily driven by the small incremental health gain relative to the additional monitoring cost, as the main modeled benefit accrued through earlier DBS treatment in only a relatively small subgroup of patients with confirmed APD.
In the probabilistic sensitivity analysis, the probability that ML-RMS was cost-effective at this threshold was 0 percent. The cost-effectiveness acceptability curve showed that the probability of cost-effectiveness exceeded 50 percent only at willingness-to-pay values above approximately €165,000 per QALY (Supplementary file 5).
Scenario analyses showed that ML-RMS could become cost-effective under assumptions generating additional downstream benefits beyond earlier APD recognition alone (Table 3). Delaying nursing home admission by 6 months in 1 percent of monitored patients reduced incremental costs to €632 per patient, but ML-RMS remained not cost-effective. When the same delay was applied to 5 percent of patients, incremental cost decreased to €113 per patient and the ICER was €25,614 per QALY gained, resulting in a positive iNMB. When nursing home admission was delayed in 10 percent of patients, ML-RMS became cost-saving and more effective, and was therefore dominant.
Base case and scenario results for ML-RMS versus standard care
Table 3. Long description
The table consists of six columns: Scenario, Incremental cost in Euros, Incremental Q A L Y s, I C E R in Euros per Q A L Y, i N M B at 35,000 Euros per Q A L Y, and Cost-effective at 35,000 Euros per Q A L Y.
* Base case: 761 cost, 0.0044 Q A L Y s, 172,023 I C E R, minus 607 i N M B, Not cost-effective.
* 1 percent nursing home admission delayed by 6 months: 632 cost, 0.0044 Q A L Y s, 142,976 I C E R, minus 478 i N M B, Not cost-effective.
* 5 percent nursing home admission delayed by 6 months: 113 cost, 0.0044 Q A L Y s, 25,614 I C E R, 41 i N M B, Yes cost-effective.
* 10 percent nursing home admission delayed by 6 months: minus 535 cost, 0.0044 Q A L Y s, Dominant I C E R, 689 i N M B, Yes cost-effective.
* 1 percent quality-of-life improvement from M L dash R M S use: 761 cost, 0.0342 Q A L Y s, 22,289 I C E R, 435 i N M B, Yes cost-effective.
* 5 percent quality-of-life improvement from M L dash R M S use: 761 cost, 0.1544 Q A L Y s, 4,926 I C E R, 4,643 i N M B, Yes cost-effective.
* 10 percent quality-of-life improvement from M L dash R M S use: 761 cost, 0.3046 Q A L Y s, 2,496 I C E R, 9,900 i N M B, Yes cost-effective.
* M L dash R M S sensitivity or specificity 95 percent slash 95 percent: 717 cost, 0.0051 Q A L Y s, 140,897 I C E R, minus 539 i N M B, Not cost-effective.
* M L dash R M S cost reduced by 50 percent: 461 cost, 0.0044 Q A L Y s, 104,289 I C E R, minus 307 i N M B, Not cost-effective.
* D B S procedure cost reduced by 50 percent: 720 cost, 0.0044 Q A L Y s, 162,753 I C E R, minus 566 i N M B, Not cost-effective.
* 12-month faster A P D confirmation and D B S assessment: 1,383 cost, 0.0216 Q A L Y s, 63,909 I C E R, minus 627 i N M B, Not cost-effective.
* M L dash R M S initiated 9 years after diagnosis: 377 cost, 0.0031 Q A L Y s, 122,322 I C E R, minus 269 i N M B, Not cost-effective.
DBS, deep brain stimulation; ICER, incremental cost-effectiveness ratio; iNMB, incremental net monetary benefit; ML-RMS, machine learning-enabled remote monitoring system; QALY, quality-adjusted life year.
Direct quality-of-life gains during active ML-RMS use had a larger effect on cost-effectiveness. A 1 percent improvement made ML-RMS cost-effective, with an incremental cost of €761, an incremental QALY gain of 0.0342, and an ICER of €22,289 per QALY gained. Larger improvements further strengthened the results. A 5 percent improvement generated 0.1544 incremental QALYs and an ICER of €4,926 per QALY gained, while a 10 percent improvement generated 0.3046 incremental QALYs and an ICER of €2,496 per QALY gained.
In contrast, scenarios focused on diagnostic accuracy, intervention costs, DBS costs, and pathway timing did not change the overall conclusion. Increasing ML-RMS sensitivity and specificity to 95 percent improved outcomes but did not make ML-RMS cost-effective. Similarly, reducing ML-RMS or DBS procedure costs by 50 percent had limited impact, with iNMB remaining negative in both scenarios. A 12-month faster APD confirmation and DBS assessment pathway increased health gains but also costs, and the ICER remained above the willingness-to-pay threshold. Delaying ML-RMS initiation until 9 years after Parkinson disease diagnosis reduced incremental costs and QALYs and remained not cost-effective.
Discussion
This early HTA suggests that ML-RMS for APD identification is unlikely to be economically favorable under base case assumptions. In the base case, ML-RMS generated a small health gain of 0.0044 QALYs per patient while increasing costs by €761, resulting in an ICER of €172,023 per QALY gained and a negative iNMB at the willingness-to-pay threshold of €35,000 per QALY. This limited value was explained by the narrow pathway through which ML-RMS was assumed to improve outcomes. The model allowed earlier recognition of APD to translate into timely DBS among eligible patients, but only 10 percent of patients with confirmed APD were assumed to receive DBS. Potential effects of ML-RMS beyond timely DBS treatment after APD confirmation, including direct quality of life benefits, improved symptom management, reduced care needs, and delayed institutionalization, were not included in the base case.
These findings indicate that diagnostic accuracy alone is unlikely to determine the value of ML-RMS in Parkinson disease. Earlier APD recognition can improve outcomes only if it initiates a clinically meaningful sequence of actions, including clinician review, specialist referral, diagnostic confirmation, treatment optimization, and appropriate initiation of advanced therapies. In the scenario analysis, increasing ML-RMS sensitivity and specificity to 95 percent improved outcomes but did not make the intervention economically favorable at the selected threshold. Although future integration of additional longitudinal data streams may alter predictive performance, the present analysis suggests that improved accuracy alone is unlikely to establish economic value unless it also translates into clinically meaningful changes in care and patient outcomes. Future validation should therefore assess not only algorithmic performance, but also whether ML-RMS changes clinical decisions and care pathways.
The potential value of earlier APD recognition is closely related to the timing and effectiveness of advanced therapies. Evidence from studies of DBS indicates that neurostimulation can improve quality of life, motor complications, and daily functioning in selected patients with Parkinson disease, including patients treated at earlier stages of motor complications (Reference Dams, Balzer-Geldsetzer and Siebert33;Reference Schuepbach, Rau and Knudsen37). However, this evidence should be interpreted carefully in the context of ML-RMS. DBS provides symptomatic benefit, but it is not established as a disease modifying therapy that slows neurodegeneration. Long-term evidence suggests that DBS may delay some disability milestones, but its effect on long-term dependency and institutional care remains uncertain (Reference Mahlknecht, Foltynie, Limousin and Poewe21).
Therefore, if ML-RMS only improves the timeliness of DBS assessment and initiation in a small subgroup of eligible patients, its economic value is likely to remain limited. A stronger value proposition would require ML-RMS to go beyond earlier APD recognition and referral for DBS assessment and to generate broader changes in the clinical trajectory. This could occur if ML-RMS accelerates specialist assessment, supports timely medication optimization, improves selection for device-aided therapy, reduces time spent with uncontrolled symptoms, enhances patient self-management, or helps prevent avoidable deterioration in daily functioning.
The scenario analyses support this interpretation. ML-RMS became economically favorable when assumed to produce benefits beyond earlier APD recognition. A 1 percent direct quality of life improvement during active ML-RMS use made the intervention favorable at the selected threshold, with an iNMB of €435. Such a benefit could arise if ML-RMS improves symptom awareness, communication with clinicians, medication adjustment, self-management, or reassurance from closer follow-up (Reference Serrano, Maita and Avila38;Reference Evers, Peeters, Bloem and Meinders39). However, ML-RMS could also introduce patient burden, anxiety, adherence challenges, and additional clinician workload (Reference Serrano, Maita and Avila38;Reference Evers, Peeters, Bloem and Meinders39). These effects should therefore be measured prospectively rather than assumed.
The largest economic impact was observed when ML-RMS was assumed to delay nursing home admission. When nursing home admission was delayed by 6 months in 5 percent of monitored patients, ML-RMS became economically favorable, although only marginally. When the same delay was assumed for 10 percent of monitored patients, ML-RMS became cost-saving and more effective. This is plausible as a value driver because institutional care is a major contributor to the economic burden of APD (Reference Kruse, Kretschmer and Lipinski40;Reference Wijers, Ravi, Evers, Tissingh and van Mastrigt41). However, direct evidence that ML-RMS or earlier APD recognition delays institutionalization remains limited. This scenario should therefore not be interpreted as an expected effect of ML-RMS. Rather, it identifies institutional care as a critical outcome for future validation.
These findings are consistent with the broader literature on digital and remote monitoring technologies in Parkinson disease. A recent HTA of remote continuous monitoring devices found that, although some devices showed diagnostic or symptom monitoring value, evidence on clinical impact was limited and cost-effectiveness was generally unfavorable under commonly applied willingness-to-pay thresholds (Reference Cox, Wade and Hodgson32). Similarly, reviews of wearable sensors and artificial intelligence-supported monitoring in Parkinson disease show that many studies focus on diagnostic accuracy, symptom measurement, or correlation with clinical scales, whereas fewer studies evaluate whether monitoring changes clinical decisions or improves patient outcomes (Reference Sapienza, Tsurkalenko and Giraitis42). The present study extends this literature by evaluating an early-stage ML-RMS intended to identify progression to APD and by showing that algorithmic performance is only one component of value.
The results illustrate the role of early HTA in guiding technology development (Reference Ijzerman and Steuten43). The purpose of this analysis was not to provide a definitive reimbursement conclusion, but to identify the circumstances under which ML-RMS could plausibly generate value, and the evidence gaps that should be addressed before full development. For ML-RMS, the key development priorities are prospective validation of diagnostic performance in the intended use population, demonstration of clinical utility, measurement of patient-relevant outcomes, and evaluation of long-term effects on advanced therapy use and institutional care.
This study has limitations. First, several model parameters were based on early evidence, extrapolation, or expert opinion because real-world data on ML-RMS performance, clinical utility, and implementation were unavailable. Second, the model assumed that ML-RMS alerts could lead to earlier APD recognition, but the pathway from alert generation to clinician review, specialist referral, diagnostic confirmation, DBS eligibility, and treatment uptake has not been prospectively validated. Third, the extrapolation assumed that all surviving patients progressed to APD by 16 years after Parkinson disease diagnosis. This simplifying assumption represented an advanced-risk cohort but may overestimate progression in more heterogeneous real-world populations. Fourth, the base case excluded direct quality-of-life effects from remote monitoring, patient reassurance, self-management benefits, monitoring burden, and alert-related anxiety. Fifth, implementation factors including adherence, workflow integration, data quality, clinician workload, alert fatigue, and recurring maintenance costs were not fully captured. Finally, the Finnish health care perspective did not fully include informal care, caregiver burden, productivity effects, or broader societal impacts relevant to APD.
Despite these limitations, this early HTA provides useful guidance for ML-RMS development. ML-RMS is unlikely to generate sufficient value if its effect is limited to earlier APD recognition and timely DBS treatment in a small eligible subgroup, but may become favorable if earlier recognition leads to meaningful changes in treatment, quality of life, or long-term care needs. Future research should prioritize prospective studies linking algorithmic outputs to clinical decisions and patient outcomes. By identifying key value drivers and uncertainties, this study supports a more targeted evidence-generation strategy.
Conclusion
This early HTA indicates that ML-RMS for timely APD identification is unlikely to be economically favorable under base-case assumptions. It could become favorable if earlier recognition leads to meaningful changes in care pathways, quality of life, or long-term care needs. Future development should address not only algorithmic accuracy, but also clinical utility, patient benefit, implementation feasibility, and effects on advanced therapy use and institutional care. These value drivers and evidence gaps illustrate how early HTA can guide ML-RMS development and evaluation in Parkinson disease.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0266462326104310.
Data availability statement
The model inputs were derived from published sources and expert input. Additional model materials are available from the corresponding author upon reasonable request.
Author contributions
C.A.G.J., W.K.D., C.A.U.-G., L.J.B., and K.D. conceived and designed the study. C.A.G.J. developed and conducted the model-based analysis. L.M. and M.K. provided clinical input on the intended care pathway and model assumptions. All authors contributed to interpretation of findings, critically revised the manuscript, and approved the final version.
Funding statement
This work was supported by the European Union’s Horizon 2020 research and innovation program through the AICCELERATE project (Grant Agreement No. 101016902).
Competing interests
The authors declare no competing interests.
Ethical standards
Because this study used published evidence, model-based analyses, and expert input and did not involve individual-level patient data or human or animal experimentation, ethics approval was not required.
Open access
