After a patient's tumor is surgically removed, doctors look for its return by drawing blood every few months and checking for fragments of tumor DNA. This circulating tumor DNA (ctDNA) is shed by surviving cancer cells, and a test called droplet digital PCR (ddPCR) can detect it. ddPCR splits a blood sample into tens of thousands of tiny droplets and counts how many light up because they hold a tumor-specific DNA fragment. Finding ctDNA this way can flag a relapse months before it appears on a CT scan. That gap is the "molecular lead time." Today the decision rule is crude. A sample is either positive or negative, or it is compared to a fixed cutoff. Two problems follow. Some patients who stay healthy produce a stray positive droplet from background noise, which leads to false alarms. Other patients shed tumor DNA only intermittently at very low levels, and a rule tuned to avoid false alarms throws that early signal away.
We propose treating each patient's series of droplet counts the way bibliometricians treat a stream of citations. The question is when the stream switches from a quiet baseline state to a sustained high-rate state, as opposed to whether any single count is high. Kleinberg's burst model does this for literature. It is a hidden-state model that charges a penalty for switching states, so isolated spikes are ignored and sustained surges are caught. The contribution is not the algorithm name. It is two changes to the detector. First, it scores raw positive-droplet counts under a Poisson measurement model with an explicit per-assay background rate. Second, it handles the irregular draw intervals that real surveillance produces. The note is equally concerned with evaluation. We want to test whether such a detector earns real extra lead time at a fixed false-alarm rate, once lead time is measured in a way that does not reward sparse CT schedules.
Core Claim
In post-resection ctDNA surveillance, a state-switching detector should do better than both "any positive droplet" and fixed allele-fraction cutoffs. The detector scores the likelihood of the actual droplet counts, including a background positive rate estimated from wild-type and no-template controls, and it penalizes transient state changes. At a matched false molecular-recurrence rate among patients who never relapse, it should reach higher sensitivity at a given horizon before radiological recurrence.
We expect most of the gain to come from the droplet-level measurement model. Only a smaller part should come from the switching penalty. A generic changepoint method equipped with the same measurement model may match the Kleinberg-style penalty. The experiments are designed to reveal which of these is true, and we treat either outcome as informative.
What Exists and What It Leaves Open
Three clinical studies motivate the proposal by showing both the value of serial ddPCR and the limits of binary calls:
- Reinert et al. followed 115 patients after resection of colorectal cancer that had spread to the liver. They used 439 serial plasma samples and tumor-informed ddPCR, with a positive/negative call per sample. They reported lead times over CT of up to 9.3 months, with a median of 2.5 months. They also flagged two problems that constrain our design. Lung metastases shed less ctDNA than liver metastases, and lead time is inflated by the three-monthly CT schedule.
- Iden et al. tracked three methylation markers by ddPCR (the TriMeth assay) in gastric and gastroesophageal junction cancer. Postoperative positivity predicted recurrence-free survival with a hazard ratio of 6.22. However, they evaluated only four discrete timepoints, and one marker (CLIP4) performed poorly.
- Chen et al. fit a static first-trimester mutant-allele-fraction cutoff of 3.15% (AUC 0.96) for progression after CAR T-cell therapy. They excluded patients with fewer than three draws.
All three reduce a time series to thresholded snapshots. We keep the full count series.
Two more studies bear on the design. Till et al. found that KRAS G12D and G12V ctDNA carry different prognostic value. This warns against pooling markers with equal weight, so our model keeps marker-specific background and shedding parameters. The Lesion Shedding Model showed that lesions within one patient shed at very different rates. It supplies realistic heterogeneity for our simulations.
The general longitudinal-biomarker literature motivates the premise that trajectories carry information snapshots miss. Salciccioli et al. found that a single neutrophil-to-lymphocyte ratio lost predictive power in sepsis. Yang et al. found an S-shaped relation between albumin change and mortality. Yang et al. also constrains the design, because it shows that incomplete serial records force large exclusions. This is why we insist on handling irregular sampling rather than dropping patients.
On the methods side, Kleinberg's model supplies the state automaton and the transition cost. Bibliometric applications such as Wen et al. and Huang et al. use it only on citation streams, where event counts are large. The research file notes that in sparse regimes the result becomes sensitive to the penalty parameter. That is exactly our regime, with around 5 to 15 draws per patient. We therefore do not hand-tune the penalty. We calibrate it to a target false-alarm rate.
Research Program
For a patient with draws at times $t_1<\dots<t_K$, draw $k$ yields $P_k$ positive droplets out of $N_k$ accepted droplets of volume $v$. Each patient has a hidden shedding state $s_k$ in {cleared, low, surging}, with state-specific concentration $\lambda_{s}$. The measurement model is
$$P_k \sim \mathrm{Binomial}\!\left(N_k,\; 1-e^{-(\lambda_{s_k} v + \beta)}\right).$$
This is the Poisson occupancy relation ddPCR already uses for quantification, as described in Lu et al.. Here $\beta$ is the per-droplet background positive rate, estimated per assay from wild-type and no-template control wells.
The detector finds the state path that minimizes
$$\sum_k -\log p(P_k \mid s_k) \;+\; \sum_k \tau(s_{k-1}, s_k).$$
The transition cost $\tau$ follows Kleinberg's form, with the penalty per upward step scaled by a constant $\gamma$. We add one extension for irregular sampling: staying in a state becomes cheaper the longer the gap between draws. The path is found exactly by dynamic programming. Multi-marker assays such as TriMeth enter as independent count likelihoods that share the hidden state. Each marker keeps its own $\beta$ and $\lambda$, so a marker that shows little signal gets little weight, as CLIP4 did in Iden et al.
The detector is run causally. At each draw it uses only past and present data, and it raises an alarm when the surging state is entered. We fit $\gamma$ once, on patients who stayed recurrence-free for the full follow-up. It is set so that no more than a chosen fraction of those patients (for example 5%) ever trigger an alarm. This turns the sparse-data tuning problem into an explicit false-alarm budget.
Lead time is computed with radiological recurrence treated as interval-censored. Recurrence is known only to fall between the last negative CT and the first positive CT, so each lead time is reported as a lower and upper bound. The primary comparison is between detectors on the same patients and the same CT schedule. That shared schedule cancels the inflation Reinert et al. warned about when detectors are compared with each other, though not in absolute lead time.
Data would come from three routes, each requiring data-sharing agreements: the raw per-draw droplet counts from the Reinert and CURE cohorts, the CAR T-cell cohort of Chen et al., and any serial ddPCR series deposited in the BloodPAC data commons. The research file lists no reported benchmark numbers on BloodPAC. The PRINCE cohort has only two draws per patient, so it is not usable for trajectories.
Because event counts are small (tens of relapses per cohort), we pair the clinical data with a simulation. It will draw lesion growth and shedding heterogeneity from the ranges in the Lesion Shedding Model paper, a lower shedding rate for lung-site recurrence, and ddPCR droplet sampling with realistic $\beta$ values. The simulation gives ground-truth onset times and allows power analysis.
Baselines
The naive reference point is the current clinical rule used by Reinert et al.: a draw is ctDNA-positive if it has at least one positive droplet above the assay's call threshold, and recurrence is declared at the first positive draw. This is the strongest existing method on the question of lead time. The file reports a median lead time of 2.5 months, up to 9.3 months, with surveillance relapse rates of 100% in ctDNA-positive versus 39.4% in ctDNA-negative patients (Reinert et al.). We will reimplement it on the same count data, with no tuning beyond the published call rule. We will also report a variant tuned to the same false-alarm budget as our method, so the comparison does not favor us through calibration alone.
The second baseline is a fixed allele-fraction cutoff chosen by ROC analysis on the training folds, in the style of Chen et al.. That paper reports AUC 0.96 for a 3.15% cutoff. However, the cutoff was for progression after CAR T-cell therapy, not post-resection recurrence, so we will refit it rather than transfer it.
The third baseline is the routine protein marker CEA tracked on the same draws. It is the cheap incumbent, and Reinert et al. report it as non-significant (HR 2.0) against ctDNA's HR 4.0.
The fourth and most dangerous baseline is a Poisson cumulative-sum (CUSUM) changepoint detector and a two-state hidden Markov model with learned transition probabilities. Both use the same droplet-level likelihood and background rate $\beta$, and both are calibrated to the same false-alarm budget. We found no reported result on Lacuna for CUSUM or HMM detectors on serial ddPCR ctDNA data.
One ablation isolates the measurement model. It keeps our switching detector but feeds it thresholded positive/negative calls or allele fractions instead of droplet counts. A second ablation removes the transition penalty, which reduces the detector to a per-draw likelihood-ratio test.
Beating the incumbent means higher sensitivity for recurrence within six months before first positive CT at the same 5% false-alarm budget among non-relapsers. The confidence interval of the paired difference must exclude zero under patient-level bootstrap. A second condition applies: no loss in median lead time.
Why not just…?
Why not just require two consecutive positive draws before calling recurrence? This is a real threat. It is a hand-built version of a switching penalty and is trivially deployable. Its weakness is that it throws away count magnitude and draw spacing. Two single-droplet positives a year apart count the same as two strong positives six weeks apart. We include it as a baseline at the matched false-alarm budget. If it ties our detector, the honest conclusion is that confirmation logic, not burst modeling, carries the benefit.
Why not just use a standard CUSUM or a fitted hidden Markov model? These are established changepoint tools, and Kleinberg's automaton is essentially a penalized hidden-state model. We expect them to perform close to our detector once they share the droplet likelihood. That is why the measurement-model ablation, not the choice of penalty, is our central test. If CUSUM matches us, we report that the gain comes from scoring counts against $\beta$, and we recommend whichever detector is simplest.
Why not just use cheaper qPCR or protein markers like CEA? The question here is the decision rule, not the platform. The platform choice also matters, because ddPCR's absolute counts and roughly 0.04% detection limit (Till et al.) are what make a count likelihood meaningful. qPCR gives a signal relative to a standard curve and has no droplet counts to model. CEA was not a significant predictor in Reinert et al. and is included as a baseline.
Why not just fit a linear slope to the ctDNA series? Most early series are zero-inflated: the draws are mostly zero droplets with occasional ones. A least-squares slope is dominated by single outliers and is undefined for patients with fewer than three draws. Chen et al. excluded such patients, and our model does not need to. We include slopes as a secondary baseline for completeness.
Proposed Experiments
The first decisive check is the simulation. With known onset times, varied shedding rates (including lung-like low shedding) and draw schedules, the droplet-count detectors should gain sensitivity at matched false-alarm rate mainly in the low-shedding regime. If they gain nothing there, the premise fails and we stop.
The second check is retrospective and uses the Reinert and CURE count data under leave-one-patient-out calibration of $\gamma$. We report five results:
- sensitivity at six months before first positive CT;
- interval-bounded lead time;
- false alarms per patient-year among non-relapsers;
- the two ablations;
- results stratified by recurrence site, to test whether the gain concentrates in the low-shedding lung subgroup Reinert et al. identified.
The third check tests marker-specific weighting on TriMeth. It asks whether modeling each marker separately recovers signal that equal-weight pooling loses to the weak CLIP4 marker.
Throughout, the outcome that would most change clinical practice is also the one we are most prepared to find. That outcome is a count-aware confirmation rule that anyone can apply on existing ddPCR output, whether or not the burst penalty itself proves necessary.
References
- Circulating tumor <scp>DNA</scp> for prognosis assessment and postoperative management after curative‐intent resection of colorectal liver metastases
- Circulating tumor DNA predicts recurrence and survival in patients with resectable gastric and gastroesophageal junction cancer
- TP53-Mutated Circulating Tumor DNA for Disease Monitoring in Lymphoma Patients after CAR T Cell Therapy
- Circulating KRAS G12D but not G12V is associated with survival in metastatic pancreatic ductal adenocarcinoma
- Lesion Shedding Model: unraveling site-specific contributions to ctDNA
- The association between the neutrophil-to-lymphocyte ratio and mortality in critical illness: an observational cohort study
- Association between albumin changes and prognosis in older patients with acute myocardial infarction
- Advances in glaucoma biomechanics from 2000 to 2024: a bibliometric study and visualization analysis
- A bibliometric analysis of artificial intelligence applied to cervical cancer
- Porcine Endogenous Retroviruses: Quantification of the Viral Copy Number for the Four Miniature Pig Breeds in China
- BloodPAC Data Commons for Liquid Biopsy Data