Women carrying twins are monitored repeatedly through the second and third trimesters, because twins carry more than ten times the preterm-birth risk of singletons (Luo et al.). At each visit, clinicians and risk tools compare a measurement or a computed risk against a single cut-off. Examples include a uterine-wall thickness below 4.3 mm or a placental hormone below 215 pg/mL. When a woman's values sit near the cut-off, ordinary measurement noise makes her alert status flip between "watch closely" and "routine" from one visit to the next. Network engineers know this failure as the ping-pong effect. They fix it with hysteresis: a rule that enters the alarm state at one threshold and leaves only after the signal recovers past a different, safer threshold, sometimes for a minimum holding time. We propose to test, on historical twin-pregnancy records, whether that asymmetric enter/exit rule reduces alert flip-flopping without delaying warnings for women who go on to deliver early. The comparison is against the simpler fixes a statistician would try first.
The contribution has two parts. The first is the alert rule and a careful comparison on a sensitivity–timeliness–stability trade-off curve. The second is an evaluation protocol for comparing dynamic alert rules on retrospective records, where clinicians' past responses to their own alarms contaminate the outcome labels. Neither part exists for twin pregnancy. The closest work, the AICare copilot (AICare), displays continuous preterm-risk trajectories for twins. It reports clinician information overload but has no mechanism that decides when an alert state should change.
Core Claim
The claim is narrow on purpose. Fix the share of early deliveries the rule flags and how far ahead it flags them. Under those constraints, an asymmetric enter/exit rule produces fewer alert state changes per patient-week than a single threshold. It also produces fewer than symmetric smoothing of the risk signal (a moving average or exponentially weighted average, which averages recent values equally in both directions) and fewer than a "debounce" rule that requires k consecutive abnormal visits.
The reason to expect an advantage is structural. Smoothing and debounce both delay entry into the alarm state. Hysteresis can keep entry immediate and put all of its conservatism on exit. That asymmetry is the one that matters clinically: a missed early warning costs a lost corticosteroid window, while a spurious "all clear" is mainly an annoyance.
If the smoothing or debounce frontier matches the hysteresis frontier, the claim is false. The proposal is built so that this outcome is reported, not hidden.
Formally, let $r_t$ be a patient's risk value at visit time $t$. This can be a measured quantity such as lower uterine segment thickness, a hormone level, or the output of a risk model. The alert enters the high state when $r_t \ge \theta_{\text{in}}$. It returns to routine only when $r_s \le \theta_{\text{out}}$ for every visit $s$ in a window of length $H$ ending at $t$. Here $\theta_{\text{out}} < \theta_{\text{in}}$ is the exit threshold and $H$ is the holding time. A single threshold is the special case $\theta_{\text{out}} = \theta_{\text{in}}$ with $H = 0$.
Where the Pieces Come From
The twin literature supplies candidate signals and their thresholds. Two single-visit studies show what these thresholds can do, and both leave the serial question open.
Başaran et al. measured lower uterine segment (LUS) thickness, the thickness of the lowest part of the uterine wall seen on abdominal ultrasound, in 58 dichorionic twin pregnancies with threatened preterm labor. A 4.3 mm cut-off predicted delivery before 34 weeks with AUC 0.707, sensitivity 71% and specificity 59%. Cervical length did not separate the groups (p = 0.060). The authors note that LUS measurements are sensitive to body mass index, bladder filling and contractions. That measurement noise is exactly what makes a single threshold flap when it is applied at every visit.
Qi et al. found that serum placental growth factor (PlGF) at or above 215 pg/mL rules out preeclampsia in twins with 98.9% negative predictive value at four weeks. PlGF is weak at confirming disease. This makes it a natural exit threshold and a poor entry threshold, which is the asymmetry hysteresis formalizes.
The networking papers supply the method. Usman et al. and Singya et al. use dual-threshold switching between optical and radio links. Singya et al. report that hysteresis bought only 0.1–0.4 dB over hard switching; its main value was suppressing switching frequency. That result constrains what we should expect. Hysteresis is unlikely to improve discrimination (AUC) and should be judged on stability at fixed sensitivity.
Lin et al. add the holding-time parameter and name its cost: route staleness. In our setting, staleness corresponds to a delayed return to routine care, which we measure directly.
Fujikawa et al. supply the clearest precedent for asymmetric timing in a deployed system. Their tunnel turns on after a 30% round-trip-time rise within six minutes and turns off only after 30 stable minutes, achieving 97% detection accuracy over three years. That supports the design. It is evidence from networks, not from obstetrics.
The retrospective-cohort literature supplies the evaluation discipline. Ye et al. used propensity-score matching (pairing treated and untreated patients with similar measured characteristics) on 932 twin pregnancies to estimate aspirin's effect. Tewari et al. used the same tool for surgery versus conservative care in prostate cancer. Both handle a single treatment decision made once at baseline.
Sternlicht et al. instead exploit variation in intervention timing. They estimate an adjusted hazard ratio of 1.018 per day of delayed intubation in COVID-19, which motivates treating alert lead time as a primary outcome. Chen et al. show that retrospective records can expose overtreatment driven by how clinicians apply thresholds. Our "alarm-weeks among term deliveries" metric is the analogue.
Research Program
The work proceeds in three stages.
Stage 1: assemble the cohort. We will build a longitudinal twin cohort from a tertiary obstetric electronic health record. The primary candidate is the BC Hospital obstetric dataset used by AICare, whose size the file does not report. If it is not accessible, we will use a comparable single-institution extract. Each pregnancy contributes every serial measurement of cervical length, LUS thickness and PlGF where available, plus routine clinical variables, chorionicity, conception mode, and time-stamped escalation events. Escalation events include admission, tocolysis (drugs to stop contractions), antenatal corticosteroids and cerclage (a stitch to keep the cervix closed). We record outcomes as spontaneous birth before 34 weeks and preeclampsia.
Stage 2: go/no-go check. Before building anything, we count how often a single threshold actually flaps on these trajectories. If fewer than a small fraction of pregnancies ever re-cross the threshold, the premise fails and we stop and report it.
Stage 3: define and evaluate rules. We build two risk signals. The first is raw LUS thickness, with the Başaran 4.3 mm value as the reference entry point. The second is a risk score learned from serial features with a simple interpretable model. On each signal we compare a single threshold, symmetric smoothing, k-of-n debounce, and hysteresis with and without holding time. A separate preeclampsia arm uses PlGF, with exit fixed at Qi's 215 pg/mL and entry tuned below it.
The evaluation protocol is the methodological part. Counting state changes is purely descriptive: it depends only on the measurement trajectory. Sensitivity and lead time are different, because they are contaminated by history. A woman flagged and treated with corticosteroids or a cerclage who then delivers at term may count as a "false alarm" only because treatment worked.
The primary analysis therefore evaluates each rule only on data before the first historical escalation. It reweights the remaining patients by the inverse probability of remaining unescalated, a correction that requires each patient to have some chance of remaining unescalated, using measured covariates. A sensitivity analysis treats "escalated then delivered at term" as an ambiguous label and reports results under both assignments. Unmeasured clinician judgment remains a threat. We will state that any claim about outcome benefit is hypothesis-generating; the stability comparison is not subject to this caveat.
Baselines
The naive reference is the single fixed threshold applied at every visit. For raw LUS this is the 4.3 mm cut-off. For the learned score it is a threshold chosen to hit a target sensitivity. For PlGF it is 215 pg/mL used for both entry and exit. This is what current single-visit evidence implies for serial use. Every other rule must beat it on state changes per patient-week at equal sensitivity, or the problem is not real.
For reference, the reported single-visit discrimination numbers are as follows. Başaran et al. report LUS at 4.3 mm with AUC 0.707 and sensitivity 71%. Qi et al. report PlGF MoM AUC 0.865, rising to 0.891 when mean arterial pressure is added. Pegios et al. (SA-SonoNet) report the cervical-length < 25 mm standard at AUC 0.673 and sensitivity 0.406, on singleton-dominated data. These are discrimination figures. No reported result was found on Lacuna for alert flapping or state-change rates in serial twin monitoring.
The strongest competitors are symmetric smoothing and k-of-n debounce. We will implement two smoothers: an exponentially weighted moving average of $r_t$ that respects irregular visit spacing, and a local-level Kalman filter (a model that tracks an underlying true level beneath noisy measurements). Debounce requires k of the last n visits above threshold to enter and k below to exit. These are what a careful statistician would deploy, and they also suppress flapping. Each competitor receives the same tuning budget as hysteresis. Each has two to three free parameters searched over an equally sized grid, chosen by nested cross-validation on a training period and reported on a later held-out period.
Hysteresis "wins" only if its frontier dominates on the held-out period at sensitivity matched within two percentage points. Dominance means fewer state changes and fewer alarm-weeks among term deliveries, without a longer median lead-time loss. Lead time is days from first alert to delivery, assessed against the seven-day corticosteroid window.
Ablations isolate the key ingredient. Asymmetric thresholds without holding time test whether the threshold gap alone matters. A holding time with equal thresholds reduces to a refractory debounce. A rate-of-change entry trigger in the style of Fujikawa tests whether LUS thinning velocity adds anything beyond its current level.
Why not just…?
Why not just keep the established single cut-off? Single cut-offs were validated for one measurement at one presentation. Başaran et al. measured LUS once at admission, and Qi et al. sampled PlGF on clinical suspicion. Nothing in that evidence says how the cut-off behaves when applied at every visit to a noisy measurement. The go/no-go flap count settles whether the single cut-off is actually a problem. If it rarely flaps, we agree with the reviewer and stop.
Why not just smooth the risk signal? This is the most serious threat. The argument against it is that symmetric smoothing delays entry as well as exit. The Sternlicht et al. timing result suggests delay is costly, but it comes from COVID-19 ventilation, not obstetrics, so it is suggestive only. The decisive evidence is the matched-sensitivity frontier comparison above. If a well-tuned exponentially weighted average or Kalman filter matches hysteresis on both state changes and lead time, we will report that smoothing suffices. We will then recommend it as the simpler tool.
Why not just require several consecutive abnormal visits? Debounce is hysteresis with zero threshold gap and symmetric delay. In sparse obstetric data with visits one to two weeks apart, requiring two consecutive abnormal visits can cost a week of warning. The ablation that keeps the holding time but removes the threshold gap measures exactly how much the asymmetry contributes.
Why not just train a sequence model to predict escalation directly? A model trained on historical escalations learns past clinician behavior, including its flapping. Records like those at BC Hospital are sparse and irregular, which AICare notes favors spurious fits. More importantly, the question is orthogonal: any learned risk score still has to be turned into an alert state. That is why we apply every switching rule on top of the learned score as well as on raw LUS.
Decisive Checks
Three outcomes would decide the proposal.
The first is the flap count under a single threshold. If it is near zero, the idea is moot.
The second is the held-out frontier comparison against the tuned smoother and debounce at matched sensitivity. It shows whether asymmetric exit buys anything that symmetric delay cannot.
The third is the PlGF arm, where the exit threshold is fixed in advance by Qi et al. rather than tuned. It tests whether hysteresis reduces state changes in a second signal without any exit-threshold tuning on our data. Agreement across the LUS arm and the PlGF arm, on a temporal hold-out and under both label assumptions for escalated-then-term pregnancies, is the minimum for recommending prospective evaluation.
References
- Secular variation and risk factors of preterm birth in Hainan, the Free Trade Port in China from 2010 to 2021
- Augmenting Clinical Decision-Making with an Interactive and Interpretable AI Copilot: A Real-World User Study with Clinicians in Nephrology and Obstetrics
- The role of the lower uterine segment thickness in predicting preterm birth in twin pregnancies presenting with threatened preterm labor
- Placental growth factor as a predictive marker of preeclampsia in twin pregnancy
- Practical Switching-Based Hybrid FSO/RF Transmission and Its Performance Analysis
- Hybrid FSO/THz-Based Backhaul Network for mmWave Terrestrial Communication
- Simple quality-of-service path first protocol and modeling analysis
- Network virtualization by differentially switched VPN for stable business communication with offshore computers
- Low-dose aspirin for primary prevention of adverse pregnancy outcomes in twin pregnancies: an observational cohort study based on propensity score matching
- Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach
- Timing to intubation COVID-19 patients: can we put it off until tomorrow?
- Practice patterns and outcomes for patients with node-negative hormone receptor-positive breast cancer and intermediate 21-gene Recurrence Scores
- Leveraging Shape and Spatial Information for Spontaneous Preterm Birth Prediction