Choosing What to Send, and in What Form, from Irregular Sensor Streams

Hospital monitors, wearables, and field sensors produce measurements at uneven times. When those measurements must cross a narrow link, such as a wireless patch, a satellite uplink, or a remote ICU dashboard, someone has to decide which of them to transmit and at what precision. Today the two relevant research communities do not talk to each other on this point. Video codecs have a mature way to make such choices: rate-distortion optimized mode selection, meaning that for each unit of data the encoder tries every available coding option and keeps the one with the lowest cost "distortion plus λ times bits". But these codecs assume frames arrive on a regular clock. Models for irregularly sampled time series handle uneven timing well, yet they treat the observation times as fixed and given. Nobody asks what a sender should transmit.

We propose a streaming encoder for irregular multivariate time series. At every observation arrival, it picks one of a small set of transmission modes: drop the observation, send only the fact and time that it happened, send a coarse value, or send a fine value. It makes this choice by simulating the receiver's own continuous-time model under each mode and measuring the resulting distortion exactly. An online controller adjusts λ so that long-run distortion stays below a target. The feature that makes this more than threshold-based transmission is the "timestamp-only" mode. Prior work shows that for clinical data, when something was measured can be more predictive than the measured value itself. That makes timing a cheap and valuable thing to transmit on its own.

Core Claim

For irregular streams, the sender's options should include sending timing information separately from values. The right way to choose among the options is closed-loop Lagrangian arbitration, in which the encoder runs a copy of the decoder, rather than a fixed error threshold. At matched bit budgets, we expect this to keep downstream task performance and reconstruction quality closer to the uncompressed stream than send-on-change thresholds, uniform binning, or learned multi-scale tokenizers.

Concretely, when observation $i$ of variable $v$ arrives at time $t_i$, the encoder evaluates each mode $m$ in a set $\mathcal{M}$ and picks

$$m_i^* = \arg\min_{m \in \mathcal{M}} \; D_i(m) + \lambda_i R_i(m),$$

where the terms are defined as follows. $R_i(m)$ is the number of bits mode $m$ costs, including entropy-coded time gaps and values. $D_i(m)$ is the distortion the receiver incurs after updating its state with what mode $m$ delivers. This distortion combines two parts: the reconstruction error on recent true values, and the divergence of the receiver's task prediction from the prediction it would make with the full stream. $\lambda_i$ is the current trade-off weight.

The trade-off weight is updated online in the style of the conformal controller of Prediction-Powered Communication With Distortion Guarantees:

$$\lambda_{i+1} = \max\big(0,\; \lambda_i - \eta\,(D_i(m_i^*) - \bar D)\big).$$

Here $\bar D$ is the target long-run distortion and $\eta$ is a step size. When realized distortion exceeds the target, λ falls and the encoder spends more bits.

Where This Sits

Three lines of work motivate the proposal.

Improving Deep Video Compression by Resolution-Adaptive Flow Coding chooses between full- and half-resolution motion fields per frame and per block by comparing Lagrangian costs. It saves up to 70% of motion bits over the DVC video codec. It supplies our decision rule, but it operates only on a fixed frame clock.

An adaptive Lagrange multiplier determination method for rate-distortion optimisation in hybrid video codecs re-estimates λ from past distortion statistics and reports up to 11% savings. Prediction-Powered Communication With Distortion Guarantees makes this kind of re-estimation carry distribution-free long-term distortion guarantees under shift. However, it does so only for discrete text tokens, with one symbol per step. We carry its controller into continuous time, where the steps are uneven and the arbitration is over several modes.

On the irregular-sampling side, two papers supply the receiver models we can simulate cheaply and causally. Latent ODEs for Irregularly-Sampled Time Series introduces the ODE-RNN, a recurrent network whose state evolves by a learned differential equation between observations and jumps at each arrival. Multi-Time Attention Networks for Irregularly Sampled Time Series introduces mTAN, which encodes continuous timestamps and attends across them. It trains more than 85 times faster than ODE-RNN and reaches 0.858 AUC on PhysioNet 2012 mortality.

Neural Controlled Differential Equations for Irregular Time Series constrains our design. Its cubic-spline control path is explicitly non-causal, so it cannot serve as the simulated receiver in a zero-delay encoder. We use it only as an offline upper reference.

Interpolation-Prediction Networks for Irregularly Sampled Time Series supplies the key motivating evidence: its observation-intensity channel alone was more predictive than the values in clinical tasks. That finding is the reason for the timestamp-only mode.

Two papers are closest in spirit. STAMP compresses irregular multivariate series with a shared multi-scale shapelet dictionary and vector quantization, but it has no per-arrival competition between modes and no online rate control. Stop&Hop learns, with reinforcement learning, when to stop and when to skip forward along an irregular timeline. It optimizes earliness of classification, not bits.

Research Program

Receiver model. The receiver is a causal continuous-time model shared bit-exactly by encoder and decoder. Our primary choice is mTAN with causal masking, because it keeps the cost of evaluating four modes per arrival low. ODE-RNN serves as a second receiver to test whether solver cost makes closed-loop search impractical, a tension the file flags explicitly.

Training for all modes. The receiver is trained on streams in which observations are randomly degraded into each mode. This teaches it to interpret drops, timestamp-only events, and coarse values. We also study a second-stage fine-tune on streams produced by the encoder itself. This matters because when the sender drops observations deliberately, the gaps themselves carry information about the data (being dropped means "close to what you predicted"), and a receiver trained only on random degradation cannot use that.

Mode set and coding. The mode set has four entries: drop, timestamp-only, coarse value, and fine value. Time gaps and quantized values are entropy coded with small learned priors that condition on the receiver state.

Myopia. The greedy per-arrival rule ignores how dropping an observation affects future costs. We quantify this loss by comparing against a learned policy trained with REINFORCE on the same Lagrangian, as Stop&Hop does for its own objective. We also compare against a one-step lookahead variant.

Guarantee. We analyze what the online λ rule guarantees when steps are uneven in time. In particular, we ask whether the long-run average bound from the text setting should be stated per observation or per unit of time. Clinical users care about the latter, and the two differ when arrival rates spike during acute events.

Baselines

Uncompressed and fixed-precision streams. The naive reference sends every observation at fixed float16 precision with its timestamp. We run it through the same receiver. It fixes the bit ceiling and the task ceiling. For the full-data task, the reported anchors are 0.858 mortality AUC on PhysioNet 2012 (mTAN) and 0.853 on MIMIC-III (IPN).

Uniform binning. Hourly binning with forward fill, then per-bin quantization, is the approach the irregular-sampling literature warns against. It is still what a practitioner would try first, so we include it at matched bits.

Tuned send-on-change thresholding. This is the strongest simple competitor, and it deserves the most care. A value is transmitted at fine precision whenever the receiver's prediction error for that variable exceeds a per-variable threshold. It uses the same receiver model and the same entropy coder. The thresholds are tuned on validation data with the same search budget we give our λ target.

STAMP. STAMP is the strongest existing learned compressor for irregular series. We apply it to trailing windows so that it runs with bounded delay, and we allow it the same delay when comparing. We report its delay separately, since our encoder has zero delay. There is no reported result on Lacuna for STAMP, or for any compressor, on PhysioNet 2012 or MIMIC-III at a bit budget. STAMP's reported numbers on SeqCombMVIrreg in the file are qualitative, apart from the symbolic baselines' blow-up past $10^{17}$. We therefore rerun it on every benchmark.

Ablations. Three ablations isolate the proposal's ingredients. Removing the timestamp-only mode tests the timing hypothesis. Replacing closed-loop distortion with an open-loop proxy (raw quantization error) tests the value of simulating the receiver. Fixing λ offline tests the online controller.

What counts as beating a baseline. We summarize the rate-distortion curve as the average bit saving at equal distortion. Beating a baseline means a lower average bit rate at equal reconstruction error, with mortality AUC no worse. The bar we set before running is at least a 10% saving over tuned thresholding at AUC within 0.005 of the uncompressed stream. It is a pre-registered target, not a prediction from prior numbers.

Why not just…?

Why not just transmit whenever the receiver's prediction error crosses a threshold? This is the real threat, and it may match us when every variable costs the same and only values matter. Our bet is that thresholding cannot express the cheap option of sending timing without values. The IPN intensity result suggests that option carries much of the task signal. Thresholding also cannot trade precision against frequency. The timestamp-only ablation against tuned thresholding settles this directly. If thresholding matches the full method once that mode is removed, and the mode itself adds nothing, we will report that simple thresholding suffices.

Why not just bin to a regular grid and reuse a standard codec? Binning erases observation times. STAMP reports grid-based symbolic coders failing catastrophically on irregular multivariate data, and IPN shows that timing is predictive. We still run binning at matched bits, so the loss is measured rather than assumed.

Why not just learn the whole policy end-to-end with reinforcement learning? A learned policy could account for future costs that the greedy rule ignores. It would, however, lose the online λ control that underpins the distortion guarantee, and Stop&Hop-style training is noisy. The REINFORCE comparison tells us how much the greedy rule leaves on the table. If the gap is large, the natural fix is to add a learned value correction inside the Lagrangian rather than abandon it.

Why not just fix λ offline? Under patient-population shift, the text experiments in Prediction-Powered Communication With Distortion Guarantees show static rate control violating its distortion target. Our shift experiment tests whether the same failure happens on physiological streams.

Proposed Experiments

We evaluate on four datasets. PhysioNet 2012 supports mortality AUC and interpolation error at held-out observations. The MIMIC-III benchmark tasks of Multitask Learning and Benchmarking with Clinical Time Series Data cover mortality, decompensation, and phenotyping. Human Activity lets us check behavior on a non-clinical sensor stream. SeqCombMVIrreg allows a direct comparison against STAMP on its own benchmark. The metrics are bits per patient-hour, reconstruction error, the gap in task AUC to the uncompressed stream, and encoder compute per arrival.

The first decisive check is the matched-bit comparison against tuned thresholding, together with the timestamp-only ablation. The second is a shift test: calibrate on one subset of PhysioNet patients, deploy on a held-out subgroup and on MIMIC-III variables shared with PhysioNet, and measure how often the per-hour distortion target is violated under online versus fixed λ. The third is a cost check. We compare mTAN and ODE-RNN receivers in the encoder loop and determine whether the closed-loop search with four modes stays within real-time budgets. The file flags the ODE-RNN as the likely failure point here.