After a narrowed leg artery is reopened with a balloon, the vessel often heals too aggressively. Smooth muscle cells and scar tissue grow inward over weeks to months and re-narrow the lumen. This regrowth is called intimal hyperplasia, and the re-narrowing it causes is called restenosis. Clinicians usually learn about it from the patient: the decision to re-treat is often triggered by a worsening symptom grade, meaning more calf pain when walking. Self-reported symptoms are an unreliable gauge of the vessel. A patient who quietly starts walking less never reaches the distance that brings on pain, and so reports feeling better while the artery closes. A patient who is anxious or depressed may report pain whose cause lies elsewhere.
We propose to treat the patient's symptom report as a second, biased observer of a hidden vessel state, rather than as the label of that state. The model has two parts trained jointly. One part infers whether the artery is re-narrowing from objective home signals: toe pulse waveforms and wrist accelerometry. The other part models whether a re-narrowed artery will actually be reported, given the patient's mood and walking behavior. This structure comes from computer vision work on human reporting bias, where the fact that people do not mention that bananas are yellow is modeled separately from whether the banana is yellow. Two things would be new: earlier, symptom-independent detection of restenosis, and a measurable estimate of which patients under-report it and why.
Core Claim
The claim is that, in post-angioplasty surveillance, much of the gap between symptom-driven endpoints and objective vessel status comes from a structured reporting process. Two forces drive it: behavioral avoidance, which suppresses reports, and affective amplification, which inflates them. A model that explicitly represents this process will detect duplex-confirmed restenosis earlier and more accurately than a model trained to predict symptom-based labels directly.
The motivating evidence comes from a randomized trial. In FREEWAY-CHINA, a trial of paclitaxel-coated versus uncoated balloons in 311 femoropopliteal patients, clinically driven re-intervention fell from 13.2% to 2.7% at 12 months. That endpoint is triggered by either Doppler stenosis of at least 70% or a worsening patient-reported Rutherford grade. Yet ankle-brachial index changes and secondary patency did not differ between arms, and the authors note that no independent core lab assessed the endpoints. The trial cannot say how much of the headline endpoint reflects vessel biology and how much reflects reporting.
A second warning concerns the clinician as observer. In intracranial stenting (E-165 Flow fraction as a predictor of intracranial in-stent restenosis), visually judged residual stenosis predicted in-stent restenosis poorly (AUC 0.59), while a computed pressure ratio did better (AUC 0.76). Visual judgments by clinicians are therefore also a weak gauge of the vessel.
The biology sets the time scale. Inflammatory recruitment of leukocytes (Simon et al.) and an early healing window (Parikh and Edelman) together imply that the vessel's fate is set in the first days to weeks. Luminal loss then accumulates over months, which is exactly the period that clinic visits sample sparsely.
Research Program
The observed symptom report at week $t$ is $y_t$, the hidden binary restenosis state is $s_t$, objective signals are $x_t$ (toe photoplethysmography, or PPG, an optical pulse waveform, plus accelerometry), and reporting context is $z_t$ (depression and anxiety scores, walking volume, how fragmented that walking is, and time since the procedure). Following the presence/relevance factorization of Yatskar et al., the model is
$$p(y_t = 1 \mid x_t, z_t) = \sum_{s \in \{0,1\}} p_v(s_t = s \mid x_{\le t})\, p_r(y_t = 1 \mid s_t = s, z_t).$$
Here $p_v$ is the vessel-state head and $p_r$ is the reporting head. The vessel-state head carries a temporal prior that $s_t$ rarely reverts from 1 to 0, which reflects the progressive nature of hyperplasia.
Two terms of the reporting head carry the clinical content. The first, $p_r(y=0 \mid s=1)$, is silent restenosis, which we expect to rise with walking avoidance. The second, $p_r(y=1 \mid s=0)$, is symptom reports without restenosis, which we expect to rise with affective distress.
Symptom reports alone cannot separate these two heads. A model fit only to $y_t$ can explain any pattern either way. We obtain identifiability by anchoring $s_t$ with duplex ultrasound at scheduled visits on all patients, and by restricting $p_r$ to depend on $z_t$ and $s_t$ but not on the waveform. The weekly symptom diaries then add information between the duplex anchors, so the model learns from far more time points than the visits alone provide.
The vessel-state head is pretrained on the UCSD toe PPG cohort, which contains 5,237 waveforms from 2,362 patients with paired ankle-brachial index measurements. In that cohort, width at half amplitude and maximum rising slope each explained about 29% of the variance in the index ($R^2$ of 0.295 and 0.292). The activity features follow Fan et al.. They showed that state-transition fragmentation and frequency-domain features from 7-day wrist accelerometry predict later pain relief (AUC up to 0.85), while concurrent pain status was predicted poorly (AUC 0.54–0.58). That pattern suggests behavior and report are distinct channels.
The new data come from a prospective cohort of femoropopliteal angioplasty patients. Each patient will contribute weekly home toe PPG, continuous wrist accelerometry in 7-day blocks each month, a weekly symptom diary (walking distance before pain and a Rutherford-style self-grade), and validated depression and anxiety scales at baseline and at 3, 6 and 12 months. Duplex ultrasound at 1, 3, 6 and 12 months defines restenosis. For sample size planning we take the 13.2% re-intervention rate in the uncoated arm of FREEWAY-CHINA as a lower bound on event frequency. Duplex-defined restenosis should be more common than re-intervention.
The design has known risks. The PPG paper discarded about 26% of signals for low quality even in the lab, so home data will be noisier. We will log signal quality and treat missing weeks as missing rather than imputing them. The wearable pulse-wave sensor has been validated only in 20 healthy volunteers, so we include it only as an exploratory secondary signal.
Baselines
The naive reference reproduces current practice. Restenosis is declared when the patient's self-reported grade worsens, scored against duplex truth at each visit. This shows how often symptoms alone are wrong and in which direction.
The strongest existing objective approach is a model trained on objective signals with duplex labels only. We use a gradient-boosted classifier on the published toe PPG feature families plus the Fan et al. accelerometry features. It ignores the diaries entirely. It uses the same patients, the same features, the same cross-validation folds by patient, and the same hyperparameter search budget as our model.
There is no reported result on Lacuna for detecting restenosis from home PPG or accelerometry. The nearest reported reference point for restenosis prediction is the AUC of 0.76 for computed flow fraction in Wang et al.. That number comes from intracranial vessels and invasive imaging, so it is context rather than a direct comparator. In the source domain, Yatskar et al. report a 3–4 point mean average precision gain from decoupling on MS COCO. That shows the factorization can help, but says nothing about clinical data.
The key ablation is collapsed supervision. We keep the same network and data but treat each diary report as a direct noisy label of $s_t$, with no reporting head. A second ablation removes affect and activity inputs from the reporting head, which tests whether the reporting structure, rather than the extra labels, does the work.
We will count the proposal as beating the baselines if both hold on held-out patients. First, it improves visit-level AUC against duplex over the duplex-only objective model. Second, it detects restenosis earlier than symptom-driven detection by a median lead time with a confidence interval excluding zero. Calibration is a secondary criterion.
Why not just…?
Why not just train on duplex labels and ignore symptoms? This is the real threat. If duplex visits are dense enough, the diaries may add nothing to detection. The comparison against the duplex-only baseline settles this directly, and we will report it whichever way it falls. Even if detection does not improve, the fitted reporting head still answers a question the duplex-only model cannot. It estimates how often restenosis is silent, and whether that silence tracks walking avoidance. That estimate bears on how trials like FREEWAY-CHINA should define endpoints.
Why not just add depression and anxiety scores as covariates in a regression on symptom reports? Adding covariates assumes mood shifts the report linearly. Our hypothesis is that bias runs in opposite directions depending on vessel state: avoidance hides true restenosis, and distress creates reports without it. A single regression on reports cannot represent that interaction without a latent state. The ablation that removes the reporting head is exactly this model with an added nonlinearity, so it tests the point.
Why not just use the ankle-brachial index or PPG thresholds at home? In FREEWAY-CHINA the index did not separate the trial arms. In the UCSD data the best single waveform feature explained under a third of the variance in the index. Single thresholds are therefore likely too weak on their own. They are included implicitly in the objective baseline, and we will also report a single-feature threshold for transparency.
Why not just use generic noisy-label learning methods? Generic noise-transition methods learn a fixed confusion matrix. The structured reporting head instead conditions on observable context, and that conditioning is what makes the bias interpretable and testable. We will include a context-free version of the reporting head as an additional ablation.
Proposed Experiments
Before enrolling anyone, we run a semi-synthetic identifiability check on the UCSD PPG data. We dichotomize the ankle-brachial index as the true state, then simulate symptom reports with known suppression and amplification rates tied to synthetic activity and affect variables. Sparse "duplex" anchors are revealed on a fraction of samples. The test is whether the model recovers the planted reporting rates and beats collapsed supervision as anchors become sparser. If it fails here, it will fail on the cohort, and we stop.
In the cohort, the primary analysis has three parts: visit-level AUC against duplex, lead time relative to symptom-triggered detection, and calibration, all compared with the baselines above.
The decisive mechanistic check uses accelerometry. Among duplex-confirmed restenoses, patients whose walking volume and active-bout length declined in the preceding month should have a higher fitted silent-restenosis probability. This is a dose-response relationship that the model was not trained to produce, and finding no such relationship would falsify the behavioral-avoidance account.
A secondary analysis compares drug-coated and uncoated balloon patients. If drug-coated balloons reduce symptom reporting more than they reduce latent restenosis, that would offer a reporting-based explanation for the gap between endpoints and objective measures seen in FREEWAY-CHINA.
References
- Twelve-Month Results From the First-in-China Prospective, Multi-Center, Randomized, Controlled Study of the FREEWAY Paclitaxel-Coated Balloon for Femoropopliteal Treatment
- E-165 Flow fraction as a predictor of intracranial in-stent restenosis
- Decreased neointimal formation in Mac-1–/– mice reveals a role for inflammation in vascular repair after angioplasty
- Tissue-Engineered Endothelial Cells Induce Sustained Vascular Healing Through Early Induction of Vascular Repair
- Seeing through the Human Reporting Bias: Visual Classifiers from Noisy Human-Centric Labels
- Photoplethysmography Signals and their Correlation with Peripheral Artery Disease
- Site-specific pain dynamics: associations between accelerometer-measured physical activity patterns and pain in older adults
- Tactile-Transparent Wearable Sensor for Clinician-Friendly Pulse Wave Velocity Monitoring and Cardiovascular Risk Profiling