When a home or ward EEG system flags a seizure, the alert is not finished at the moment of detection. A clinician usually has to see a short clip of the raw recording to confirm the event before acting, because automated detectors still raise false alarms. That clip is a short, time-limited file transfer. It travels over a home router or cellular link that is shared with video streams, downloads, and other devices, and each of those runs its own congestion control, the algorithm that decides how fast a sender pushes data into the network. Whether the clip arrives in time therefore depends on three things at once: the patient's detector, the network path, and the mix of competing senders. Published evaluations almost always fix two of these and vary one, and often fix all three: one patient, one trace, one competitor. This is the networking and clinical version of the single case study. It shows that a system can work. It does not show that the result holds as a pattern.
This note proposes a benchmark and a measurement study built around one quantity: the probability that a confirmable seizure alert reaches the clinician within a deadline. We measure it across every combination of patient, path, and competing traffic mix, so the study can do two things. It can locate where deadline misses actually come from. It can also test a methodological claim: choosing a mitigation from a single patient, path, or competitor often picks the wrong one. To test the claim we report how often rankings reverse, and how many evaluation conditions are needed before a ranking stops changing.
Core Claim
Our hypothesis is that the tail of end-to-end time-to-confirmation is driven mainly by cross-protocol competition at the bottleneck, and less by the base path or the detector's own latency. Cross-protocol competition means different congestion-control schemes fighting over the same buffer. We expect the ordering of standard mitigations to reverse between single evaluation conditions often enough that single-condition studies are unreliable. The mitigations are plain slow start, history-based "hot start," priority marking, queue management at the router, and sending less data.
Three strands of prior work motivate the claim. Pantheon built a global testbed and calibrated emulators. It found that no protocol wins everywhere, which is direct evidence against single-environment evaluation. But it measures only flows of the same scheme competing with each other, and only long "full-throttle" transfers, and it names cross-scheme competition as future work. Our deadline-bound short clip competing with different schemes sits exactly in that stated gap. On the long-run behavior of equation-based rate control shows that a sender tuned to match TCP's average rate still takes an unfair share, because its smooth traffic changes how the shared buffer drops packets. It follows that fairness and latency under mixed traffic have to be measured, not derived. WiseStart reuses cached path history to start transfers with a large window. That makes short transfers much faster. The paper tested fairness only against Cubic, which leaves open whether a clip sent aggressively harms, or is harmed by, other schemes.
The clinical literature plays a different role. It constrains how outcomes should be scored, and it does not supply a timescale. The cohort studies find threshold-shaped harms from delay: spasm duration over 3 months before response carried a relapse hazard ratio of 3.11 in the infantile spasms cohort, and immunotherapy after 28 days carried an adjusted odds ratio of 6.432 for chronic epilepsy in the autoimmune encephalitis cohort. These windows span days to months, not seconds, so we borrow only their shape. The same papers also note that turning continuous delay into a yes/no cutoff hides the underlying dynamics. For that reason we sweep the deadline $D$ and report full delay distributions, rather than defending any single clinical cutoff, which the file does not supply for seizure alerts.
Research Program
We define the end-to-end latency for a true seizure $s$ as $T_s = t^{\text{det}}_s + t^{\text{xfer}}_s$. Here $t^{\text{det}}_s$ is the time from annotated onset to detector firing, and $t^{\text{xfer}}_s$ is the time for the alert plus a fixed-length raw clip to arrive completely at a remote endpoint. The primary metric is the deadline-hit curve
$$H(D) = \Pr\left[\,T_s \le D \text{ and } s \text{ detected}\,\right],$$
computed over seizures in a given condition and swept over $D$. Missed detections count as misses at every $D$. False alarms do not enter $H$. They are replayed as extra clip transfers, because they add load and cost clinicians attention.
On the detector side we use the CHB-MIT scalp EEG database: 916 hours from 24 patients with 173 labeled seizures. We reimplement the patient-specific SVM from Application of Machine Learning To Epileptic Seizure Detection. It uses 2-second windows across 8 frequency bands and 18 channels, with three windows stacked. We use it because its detection latency and false-alarm timing are documented per seizure.
On the network side, each detected event, true or false, triggers a clip upload. The upload runs through Mahimahi's delay and link shells (Mahimahi), driven either by the 70 mobile traces used by WiseStart or by Pantheon-calibrated emulators. Pantheon emulators are specified by bottleneck rate, delay, queue size, loss, and delivery process. Home neurotechnology is a realistic setting for such links: the T15 intracortical interface averaged 10 hours of daily at-home use.
Competing traffic is drawn from a fixed menu of mixes: Cubic bulk flows, a BBR flow, an equation-based smooth-rate flow, an aggressive UDP bulk sender in the style of UDT, and a two-subflow multipath flow with coupled CUBIC. The last stands in for a phone using Wi-Fi and cellular at once.
The analysis has two parts. The first is a variance decomposition of $\log T_s$, fitted with random effects for patient, path, and competitor mix and their interactions. It tells us which factor dominates the tail. The second targets the single-case question directly. For every single cell (one patient × one path × one mix), we rank the mitigations and compute Kendall's rank correlation against the ranking on the full cross-product. We then bootstrap how many randomly sampled cells are needed before the ranking agrees with the pooled one 95% of the time. That number is our quantitative answer to "how many cases make a pattern" for this problem.
Alongside the benchmark we test one simple mechanism that a skeptic cannot get for free from existing tools: progressive clip encoding. The sender first transmits a coarse version of the clip, made of the channels the detector weighted most, heavily downsampled. It then fills in full resolution. The idea is that something confirmable arrives early even when the full clip would miss $D$. Whether a coarse clip is actually confirmable is a clinical judgment we cannot settle here. We therefore report it as a separate curve $H_{\text{coarse}}(D)$ and do not merge it into $H$.
Baselines
The naive reference is the unmodified default. The clip is sent over Cubic with standard cold-start slow start, the bottleneck uses a drop-tail buffer, and there is no traffic marking. Every other method is judged by how much it raises $H(D)$ over this default at fixed $D$, and at what cost to competing flows' throughput.
The strongest existing sender-side method is WiseStart (Cold Start or Hot Start?). It caches bottleneck bandwidth and propagation delay per client, starts with a large window paced at twice the estimated rate, and has an explicit drain phase. It is the strongest competitor because the clip is precisely a short, deadline-bound transfer, and WiseStart is built for short transfers. Its paper reports a 25.43% reduction in average request completion time at connection establishment and a 52.34% reduction in P99 latency in production. We run it on identical traces, patients, and competitor mixes. Its history cache is warmed on the same number of prior transfers per path as any other history-using method. To beat it, a method must show a higher $H(D)$ at the median deadline, with a 95% bootstrap interval excluding zero, in the competitor mixes where WiseStart was never tested, meaning everything except Cubic. It must also impose no worse a throughput loss on those competitors.
Two network-side baselines complete the comparison. The first is strict priority marking of clip packets at the bottleneck. The second is per-flow fair queueing with delay-targeted active queue management at the bottleneck. We found no reported result on Lacuna for either on this question. Both are configured with default parameters, with no per-condition tuning for any method.
The key ablation is the evaluation itself. We compute each method's apparent benefit using only the single most-cited style of setup, which is one patient, one stationary trace, and one Cubic competitor. We then compare that with its benefit on the full cross-product. A second ablation removes progressive encoding from our sender to isolate its effect. Pantheon's emulators reproduce real-path throughput and delay with an average error of 17%, so we treat method differences smaller than that on emulated paths as unresolved.
Why not just…?
Why not just run detection on the device and send only a tiny alert? For fully automated response, this is the right answer, and the benchmark includes it as a condition: a delivered alert with no clip. The difficulty is confirmation. The generic "Reveal" detector on CHB-MIT caught 61% of seizures with 33 false alarms per day, versus 96% and a median of 2 per day for patient-specific models (Application of Machine Learning To Epileptic Seizure Detection). With false alarms at those rates, a human usually needs raw signal. If the experiments show that clip transfer rarely changes $H(D)$ over the plausible range of $D$, the study's conclusion is that the network does not matter here. That would be a publishable negative result.
Why not just deploy fair queueing with active queue management at the router? This is the most serious threat. If it restores $H(D)$ across all competitor mixes, the practical recommendation becomes "fix the bottleneck." The catch is that patients rarely control their home router or their cellular carrier's queue. We therefore report results separately for bottlenecks with and without such queueing. The sender-side question matters only for the second group, and we will say so plainly if that group turns out to be small in effect.
Why not just mark clip packets high priority? Marking helps only where the bottleneck honors it, which it often does not on consumer paths, and a mark cannot compensate for a misestimated link rate on volatile paths. Strict priority can also starve other traffic during bursts. The experiments measure both effects directly: $H(D)$ with and without the mark being honored, and the throughput lost by competing flows.
Why not just use WiseStart as is? WiseStart may well be sufficient. It is our strongest baseline for that reason. The open question is whether its fairness and its latency gains hold against BBR, smooth-rate, aggressive UDP, and multipath competitors, which its evaluation did not include. The cross-product answers that question.
Proposed Experiments and Decisive Checks
The main experiment runs all methods over 24 patients × about 70 traces plus a grid of Pantheon-calibrated emulator settings × 6 competitor mixes × 2 bottleneck queueing policies. Deadlines are swept from a few seconds to several minutes. Every result is reported as a full $H(D)$ curve with per-factor breakdowns.
Three outcomes would be decisive:
- Tail source. If competitor mix explains less of the variance in $\log T_s$ than the path does, the core claim about cross-protocol competition is wrong.
- Single-case unreliability. If single-cell rankings agree with the pooled ranking (rank correlation near 1 in most cells), the single-case critique does not apply to this problem, and a small evaluation is justified.
- Mitigation gap. If fair queueing closes the gap everywhere it is present, and WiseStart closes it everywhere else without harming competitors, then no new sender is needed. The contribution reduces to the benchmark and its minimum-evaluation-size curve, and we would report it that way.
References
- Pantheon: the training ground for Internet congestion-control research
- On the long-run behavior of equation-based rate control
- Cold Start or Hot Start? Robust Slow Start in Congestion Control with A Priori Knowledge for Mobile Web Services
- Epileptic spasms relapse is associated with response latency but not conventional attributes of post‐treatment <scp>EEG</scp>
- Acute Symptomatic Seizures and Risk of Epilepsy in Autoimmune Encephalitis: A Retrospective Cohort Study
- Application of Machine Learning To Epileptic Seizure Detection
- Mahimahi: Accurate Record-and-Replay for HTTP
- A Personalized and Adaptable User Interface for a Speech and Cursor Brain-Computer Interface
- Teraflows over Gigabit WANs with UDT
- Coupled CUBIC Congestion Control for MPTCP in Broadband Networks