The Measurement Problem Behind Two Decades of Neurology Trial Failures

On April 28, the FDA announced the successful initiation of two real-time clinical trials and released a Request for Information on a broader pilot program for AI-enabled early-phase trials. The agency now treats continuous, real-time data sharing as a regulatory direction of travel rather than a theoretical future. The neurology field should pay close attention. No therapeutic area has more to gain from the shift, and none has been worse served by the model it replaces. 

The numbers are familiar but worth restating. Roughly 90% of drug candidates that enter clinical development never reach patients. Phase 3 readouts for Parkinson’s and Alzheimer’s have produced a near-unbroken string of failed primary endpoints over the last decade. A successful drug program now routinely costs more than a billion dollars and a decade of work. Even the rare approvals, with donanemab in Alzheimer’s being the most visible recent example, often struggle to translate into real-world uptake. Parkinson’s has seen meaningful advances in symptomatic treatment, but the field still lacks an approved disease-modifying therapy that slows or halts the underlying disease process. That gap has held for decades. 

It is worth being humble about what AI will and will not do in this field. The honest answer is that no one knows the long-term trajectory. What the industry can say with reasonable confidence is that AI is not going to replace clinical trials as a regulatory mechanism for drug approval in the next ten years. The FDA’s RTCT initiative is not pointing in that direction either. It is pointing at something narrower and more achievable: helping sponsors learn earlier which patients are most likely to progress, which signals are clinically meaningful, and whether a therapy is showing enough evidence of effect to justify the next stage of investment. That is the question neurology sponsors should be planning around, because that is what regulators are signaling they will support. The opportunity for the field is to take the endpoints sponsors and regulators already trust and capture them with the sensitivity, frequency, and ecological validity that current clinic-based methods cannot deliver. 

Why Phase 3 in Parkinson’s is so long 

The standard motor endpoints in Parkinson’s, including patient-reported OFF time, MDS UPDRS Part II progression, and fall frequency, are clinically meaningful and regulator-accepted. They are also captured today through episodic clinic visits and paper diaries, which makes them both noisy and slow-moving. UPDRS typically progresses only a few points per year in untreated PD, and the within-patient measurement variability on a given visit is roughly the same magnitude. A year of true disease progression can therefore sit inside the noise floor of the instrument used to detect it. The consequence is well-known to anyone who has designed a Phase 3 in PD: trials have to be both long enough to accumulate enough signal above the noise and large enough to average that noise out across patients. The 18-to 24-month, multi-hundred-patient PD Phase 3 is the answer the field has been forced into by its measurement infrastructure. Faster, more sensitive measurement would change what an adequately powered PD trial has to look like. 

Disease expression in Parkinson’s varies from patient to patient and within the same patient over the course of a single day. Clinic visits scheduled weeks or months apart miss most of that variation by design, while continuous measurement between visits surfaces it. The distinction matters across three steps that are too often collapsed into one. Wearables and digital measures generate real-world data. Validated analytics turn that data into interpretable disease signals. Fit-for-purpose study design turns those signals into real-world evidence that can support regulatory decision-making. Simulation, layered on top, lets sponsors pressure-test trial design, patient selection, and go/no-go decisions before committing years and hundreds of millions of dollars. Each layer matters, and conflating them is how the field loses credibility with regulators. 

That means tightening confidence intervals on OFF time, detecting MDS UPDRS Part II inflections earlier, and surfacing fall events the patient never reported. Smaller cohorts, shorter durations, and earlier go/no-go decisions follow from better measurement of accepted endpoints. 

From diagnostic label to responder population 

The clinical heterogeneity of Parkinson’s is well-accepted. Tremor-dominant patients, postural-instability-and-gait-disorder patients, cognitive-predominant patients, and several proposed molecular sub-types all behave differently and likely respond to therapy differently. Broad diagnostic labels mask biologically distinct patient groups. The standard trial design, which enrolls a heterogeneous PD population, measures a population-average endpoint, and hopes for an effect, is structurally incapable of finding therapies that work for a subset. 

Neurology will not copy oncology overnight. Companion diagnostics in oncology rest on decades of tumor biology, biomarker validation, and regulatory precedent that neurology does not yet have. But oncology has shown what becomes possible when trial design, diagnostics, and treatment selection are built around biological subtypes rather than broad disease labels. The path forward in neurology has two converging substrates. On the front end, genomic foundation models are beginning to make variant-effect prediction tractable for the Parkinson’s-relevant gene families, including GBA, LRRK2, SNCA, and ASNS. That capability is the substrate any precision neurology program will eventually require. On the back end, continuous digital phenotyping provides the longitudinal response signal that lets sponsors identify who got better and by how much. 

Consider a Parkinson’s therapy that appears ineffective across a broad trial population. In a traditional trial, that is the end of the program. With continuous endpoint measurement and biologically grounded subtyping, sponsors may discover that the drug benefits a specific group: patients with a particular genetic profile, motor phenotype, or progression trajectory. Whether a failed program becomes an approved precision therapy depends on whether the trial was instrumented to find that group in the first place. 

Three questions before the protocol locks 

The pragmatic question for a clinical development team is whether their next Phase 2 in PD or Alzheimer’s is instrumented to detect a real effect if one is there. Three concrete questions are worth asking before the protocol locks. 

Is the trial capturing the accepted primary endpoint continuously, or only at scheduled visits? Is the enrollment strategy stratifying on genetic and phenotypic sub-type, or assuming a uniform responder population? Is the dataset structured so that a negative primary endpoint can still yield a defensible responder analysis rather than a write-off? 

None of this requires regulators to accept novel endpoints. It requires sponsors to take the endpoints they already have and measure them properly, which is the direction the FDA’s RTCT initiative is pointing. The neurology field should weigh in. 

What the first real win looks like 

Within the next several years, a Parkinson’s patient will receive a therapy whose responder population was identified through genomic stratification and validated through continuous digital phenotyping in the registrational trial. That program ends the decades-long drought of disease-modifying PD approvals. The technology to run it exists now. The remaining question is which sponsors decide to build it into their next protocol before another wave of Phase 3 failures forces the change.

Photo: ClaudioVentrella, Getty Images

This post appears through the MedCity Influencers program. Anyone can publish their perspective on business and innovation in healthcare on MedCity News through MedCity Influencers. Click here to find out how.