| Journal of Biomedical Systems and Engineering

Received: 24 July 2026; Revised: 06 September 2026; Accepted: 09 September 2026; Published Online: 21 September 2026.

J. Biomed. Syst. Eng., 2026, 1(1), 26801 | Volume 1 Issue 1 (September 2026) | DOI: https://doi.org/10.64189/bse.26801

This article is licensed under Creative Commons Attribution NonCommercial 4.0 International (CC-BY-NC 4.0)

Compression-Aware Vital-Sign Anomaly Detection: A

Synthetic Data Ablation of Pruning, Quantization, and

Knowledge Distillation, With A Protocol for Clinical

and Edge-Hardware Extension

Pushpak Basak* and Sureeti Sahu

Department of Computer Application, Maharaja Agrasen International College, Raipur, Chhattisgarh, 492001, India

*Email: pushpakbasak49@gmail.com (Pushpak Basak)

Abstract

Continuous vital-sign monitoring on wearable and bedside healthcare IoT devices requires machine-learning

models small enough for microcontroller- or single-board-class hardware. Published compression studies

rarely combine a clinically motivated anomaly detection task with physical edge-hardware evaluation; a

nonexhaustive literature search did not locate such a study. In this paper, pruning/quantization/distillation

ablation is performed on a synthetic vital-sign proof-of-concept dataset, and the additional protocol needed to

extend the study to real clinical data and physical hardware is specified; the executed findings are distinguished

from the proposed extension work. A synthetic multivariate vital-sign generator produced 300 patients

contributing 2,400 windows in total (8 windows per patient; heart rate, respiratory rate, SpO2, systolic blood

pressure, and temperature), with a patient-level 210/45/45 train/validation/test split and NEWS2-based

threshold labeling. A CNN-LSTM baseline was trained and evaluated in PyTorch on a CPU against five

alternatives: structured channel pruning, partial dynamic posttraining 8-bit integer quantization (linear/LSTM

layers only), knowledge distillation to a 1,662-parameter GRU student, and two combinations of the above. The

baseline reached F1 = 0.955 (AUROC 0.993) on the synthetic test set. Structured pruning of the convolutional

front end left the whole-model size essentially unchanged because the LSTM back end holds 93.3% of the

parameters; the global sparsity reached only 0.44–2.65% across the five pruning levels. Knowledge distillation

to the GRU student reached F1 = 0.912, a 40.8-fold reduction in multiply accumulate operations, against F1 =

0.819 for the identical architecture trained without distillation. Partial dynamic quantization reduced the

serialized size by 66.2%, with essentially unchanged performance, although the host-CPU latency increased

because of unoptimized quantization kernels. No power, energy, or edge-hardware measurements are reported

or estimated. These findings characterize a synthetic-data, CPU-sandbox compression study; they should not be

read as evidence of clinical effectiveness, energy efficiency, or real-time performance on target hardware, for

which a self-contained extension protocol is specified separately.

Keywords: Model Compression; Structured Pruning; Knowledge Distillation; Posttraining Quantization; TinyML;

Healthcare IoT; Reproducibility.

1. Introduction

Continuous, low-power vital-sign monitoring is important for healthcare internet-of-Things (IoT) devices, and

performing inference directly on miniaturized edge hardware is attractive from the perspectives of both privacy

and bandwidth. Executing modern deep-learning models on such constrained platforms typically requires

model compression, most commonly through pruning, quantization, or knowledge distillation. Survey evidence

from the TinyML compression literature indicates that the benefit of pruning and distillation is task dependent,

whereas quantization is consistently valuable under memory- and power-constrained conditions, motivating

an evaluation of all three techniques together rather than in isolation.

[1]

Existing studies on model compression

for physiological time series typically evaluate a single technique, or a fixed pipeline, on a task chosen for

convenience rather than clinical relevance and rarely report which architectural component of the model is

responsible for the achieved compression. Reviews of early warning systems based on vital signs and cancer-

specific deterioration scores establish that recurrent architectures are the standard choice for this task, and

surveys of TinyML in wearable healthcare together with structured-pruning research specific to recurrent

networks confirm that pruning strategies effective for convolutional networks do not transfer automatically to

LSTM- or GRU-dominated models.

[2-5]

What is missing from this literature is a controlled ablation, split at the

patient level, that evaluates pruning, quantization, and knowledge distillation side by side on a clinically

grounded vital-sign anomaly detection task, reports which architectural component drives the achieved (or

unachieved) compression, and states plainly which claims are supported by the executed experiments rather

than by a proposed extension.

The specific problem addressed here is how the accuracy, size, and latency trade-offs of structured pruning,

posttraining quantization, and knowledge distillation compare for a CNN-LSTM vital-sign anomaly detector

when evaluation is constrained to a generic CPU sandbox rather than physical edge hardware. This constraint

reflects a common practical situation in early-stage compression research: the compression pipeline and

ablation protocol can be fully developed and evaluated before physical target hardware or credentialed clinical

data become available, and the resulting synthetic data and CPU-sandbox findings need to be reported honestly

as such rather than extrapolated into claims that the evidence does not support. This paper provides a

controlled evaluation, run only on a CPU, of how pruning, quantization, and knowledge distillation trade

accuracy for compression under a fixed latency budget on a clinically grounded synthetic vital-sign detection

task. Specifically, it (i) introduces a synthetic vital-sign benchmark, split at the patient level and controlled for

leakage; (ii) implements and executes an ablation across five compression configurations and two combinations

thereof; (iii) identifies and explains a negative result such that convolutional-layer pruning does not

meaningfully compress an LSTM-dominated architecture, with a direct connection to the structured-pruning

literature; (iv) demonstrates, through a matched ablation, that knowledge distillation recovers most of the

accuracy of a student model 40 times smaller; and (v) specifies a self-contained protocol (Appendix A) for

extending this study to real clinical data and physical edge hardware without modifying the software already

developed.

2. Related work

MIMIC-III is the largest publicly available ICU database used to predict deterioration from vital signs and to

design early-warning systems.

[6]

A systematic scoping review of early warning systems based on machine

learning confirms that LSTM- and attention-based encoder–decoder architectures are the dominant choice for

this task, and an early warning score based on deep learning, developed for oncology patients, illustrates the

clinical stakes of building such systems well.

[2,3]

The literature search conducted for this manuscript did not

identify prior work combining MIMIC-based NEWS2 labeling with controlled pruning/quantization/distillation

ablation and direct per-device energy measurement. Benchmarking studies on the edge, such as

DeepEdgeBench, report inference time and power for a vision model across a Raspberry Pi 4, Jetson Nano, and

other single-board computers, finding that device rankings vary with inference rate and accelerator use rather

than following a fixed order.

[7]

The protocol specified in Appendix A of this work adopts a comparable

multidevice, multirate measurement approach for the vital sign model rather than assuming a fixed device

ranking in advance.

Structured-pruning surveys of convolutional networks report that channel- or filter-level pruning accelerates

inference only when the pruned dimension holds a substantial share of a layer's parameters or FLOPs; this is

precisely why pruning the CNN front end in Section 6.3 produced no meaningful compression, since the CNN-

LSTM architecture's parameters and computation budget are concentrated in its recurrent layers.

[8]

Structured

pruning specific to recurrent networks shows that meaningful compression of LSTM- or GRU-based models

instead requires pruning strategies applied to the recurrent weight matrices themselves or dedicated low-rank

factorization of the recurrent layers.

[5]

This is a direction not pursued in the present execution but identified as

a priority in Section 8.2. TinyML reviews of wearable and ambulatory healthcare devices report that memory

and energy constraints, not the choice of compression method alone, dominate the feasibility of on-device

deployment, reinforcing the importance of reporting size, MACs, and latency separately rather than treating

compression as a single scalar quantity.

[4]

A study combining quantization and knowledge distillation for a

nonmedical indoor localization task demonstrates that combining these two techniques can meet the memory

budgets that neither technique achieves alone; although that work's application domain differs from the

present one, its methodological finding is consistent with the distillation ablation in Section 6.5, which

outperforms the pruning ablation in Section 6.3.

[9]

The application of knowledge distillation specifically to

wearable single-lead ECG arrhythmia classification and structured pruning combined with binarization for

sleep apnea detection from wearable IoT sensors resulted in substantial compression with limited accuracy

loss for clinical time series tasks, supporting distillation and quantization as the more promising directions for

this architecture family.

[10,11]

The teacher‒student framework used for distillation in this work (Section 4.3,

Equation 9) follows the original formulation of Hinton et al.

[12]

The broader motivation for pursuing model

compression and reducing the energy and computational footprint of machine-learning systems is consistent

with the Green AI research agenda; however, as stated throughout this manuscript, no energy measurements

were performed in the present execution.

[13]

3. Contributions

Building on the gap identified in Section 1, the contributions of this execution are as follows:

• A synthetic vital-sign benchmark, split at the patient level and controlled for leakage, with NEWS2-

grounded threshold labeling, is usable as a drop-in substitute until MIMIC access becomes available

(Section 4.1; Appendix A, Protocol A).

[14]

• A working compression pipeline supports ablation across structured pruning, posttraining dynamic

quantization, knowledge distillation, and combinations of these techniques (Section 4.3).

• A negative result with practical implications: Convolutional-layer pruning has a negligible effect on the

whole-model size for this LSTM-dominated architecture, which is consistent with the recurrent-

network pruning literature (Section 6.3).

[5]

• A positive result with practical implications: Knowledge distillation recovers most of the accuracy gap

between a 1,662-parameter student and its nondistilled counterpart of identical architecture (Section

6.5).

• A self-contained protocol (Appendix A) for the MIMIC-based and physical-hardware extension of this

work, requiring no modification to the software already developed, was used.

4. Materials and methods

4.1 Synthetic dataset

The synthetic-data generator produces 300 patients, each contributing 8 windows of 60 timesteps across 5

channels (heart rate, respiratory rate, oxygen saturation, systolic blood pressure, temperature), for 2,400

windows in total. All channels for a given patient are generated as a first-order autoregressive (AR(1)) drift

around a patient-specific baseline within the NEWS2 normal band for that channel.

[14]

A subset of windows

(27.0% overall: 27.3% train, 27.2% validation, 25.3% test) contains an inserted event of random duration (15–

35 timesteps) in which one or two channels are driven, via a smooth ramp rather than a step change, to a value

45–85% beyond the width of the NEWS2 normal band for that channel. The label is therefore a deterministic

function of the NEWS2-band threshold crossing on any channel, not an arbitrary statistical anomaly score. The

exact AR (1) coefficient and noise standard deviation used by the generator are not restated in this manuscript

beyond the ranges above; this is noted as an outstanding reproducibility item in Section 10.

The patient-level split is 210/45/45 train/validation/test, with no patient appearing in more than one split,

resulting in 1,680/360/360 windows (Table 1). Normalization parameters (zero mean, unit variance) were fit

on the training set only and then applied to all the splits, avoiding normalization leakage. The positive-class rate

is close to the generator's 27.0% target across the training and validation splits (27.3% and 27.2%) but

somewhat lower on the test split (25.3%), with only 45 patients contributing to the test set; this two-point gap

is consistent with ordinary sampling variability at that sample size, although no formal significance test was

performed, and this should be read as a plausible explanation rather than a confirmed one. The generator is not

a substitute for MIMIC-III/IV or WESAD: it does not model irregular sampling, missingness, cross-channel

correlation from comorbidities, or realistic class imbalance, and all detection metrics reported in this

manuscript characterize a synthetic-data proof-of-concept rather than clinical performance. Appendix A,

Protocol A, specifies how the synthetic generator can be replaced with a real MIMIC-derived extraction while

preserving the same (features, labels, patient ID) interface, requiring no changes to the downstream training or

evaluation code.

Table 1: Synthetic dataset characteristics

Property

Value

Patients (total)

300

Windows (total)

2,400 (8 per patient)

Window length/channels

60 timesteps × 5 channels (HR, RR, SpO2, SBP, Temp)

Train/val/test patients

210/45/45 (patient-level, no overlap)

Train/val/test windows

1,680/360/360

Positive-class rate (train/val/test)

27.3%/27.2%/25.3%

Labeling rule

NEWS2-band threshold crossing 1–2 affected channels per positive window

Data type

Synthetic proof-of-concept (not MIMIC-III/IV, not WESAD)

4.2 Model architecture

Baseline/teacher (CNN-LSTM): Two 1D convolutional layers (32 channels each, kernel sizes 5 and 3, each

followed by batch normalization and ReLU), a two-layer LSTM (64 hidden units per layer, dropout 0.2 between

layers), and a linear classification head (dropout 0.3 before the head).

Student (LightGRU): Single-layer GRU with 20 hidden units and a linear head, used as both a knowledge-

distillation target and a lightweight deployment candidate.

The system and model architecture are illustrated in Fig. 1 and detailed description is provided in Appendix C.

The baseline parameter count, computed directly from the architecture definition, is 62,562 total parameters

(258,075 bytes = 252.0 KB stored as FP32): 896 for the first convolution and batch norm, 3168 for the second,

58,368 for the two-layer LSTM, and 130 for the fully connected head. The estimated MAC count for a 60-time

step input, computed directly from each layer’s definition rather than a profiler, is 3,673,088 MACES: 48,000

MAC for the first convolutional layer, 184,320 MAC for the second, 3,440,640 MAC for the LSTM, and 128 MAC

for the fully connected head. The LSTM layer therefore comprises 93.3% of the parameters and 93.7% of the

MACs (Fig. 2, Table 2), which explains the negative pruning result reported in Section 6.3. An earlier draft of

this manuscript reported this figure as "over 98%", a value that could not be reproduced from the architecture

definition and has been corrected to 93.3% throughout. The training was performed with the Adam optimizer,

a learning rate of 1×10⁻³ (5×10⁻⁴ for postpruning fine-tuning), a batch size of 32, up to 40 epochs with early

stopping (on the validation F1, patience 10), gradient-norm clipping of 2.0, and a fixed random seed (42).

Fig. 1: System and model architecture

Fig. 2: (a) Parameter distribution and (b) MAC distribution across CNN-LSTM submodules, computed by direct per-

layer accounting

Table 2: Baseline CNN-LSTM parameter and MAC distribution by submodule

Submodule

Parameters

% of total

MACs/window

% of total

Conv1 + BatchNorm1

896

1.4%

48,000

1.3%

Conv2 + BatchNorm2

3,168

5.1%

184,320

5.0%

LSTM (2 layers, 64 hidden)

58,368

93.3%

3,440,640

93.7%

Fully connected head

130

0.2%

128

0.0%

Total

62,562

100%

3,673,088

100%

4.3 Compression configurations

Table 3 lists every configuration in the underlying study plan together with its execution status. Configurations

marked “not executed” are not estimated or approximated; each depends on a resource unavailable in this

environment (a locked deployment runtime for quantization-aware training or the output of an earlier, not-yet-

executed step).

Table 3: Compression configurations: identifier, description, and execution status

Configuration

Status

Baseline CNN-LSTM, no compression

Executed

Structured (L2, channel-level) pruning of conv1/conv2 at

10/20/30/40/50%, with postprune fine-tuning

Executed (5 sparsity levels)

Partial dynamic posttraining INT8 quantization (Linear +

LSTM modules)

Executed, partial scope (Conv1d unsupported by

PyTorch's dynamic-quantization path)

Quantization-aware training (QAT)

Not executed. Requires a locked deployment runtime

(TFLite/TensorRT); see Appendix A

Knowledge distillation, CNN-LSTM teacher to GRU student

(T = 4.0, α = 0.5), plus a matched no-KD ablation

Executed

Structured pruning (30%) of the teacher, then KD to the

GRU student

Executed

PTQ + KD

Not executed this pass

Pruning + QAT + KD (full combination)

Not executed. Depending on M3

Structured pruning uses channel-level L2-norm structured pruning on both convolutional layers, followed by

mask fixation and 8 epochs of fine-tuning at a reduced learning rate. The pruning implementation zeroes

channel weights without physically shrinking the underlying tensor unless the layer is explicitly rebuilt with

fewer channels; this is a sparsification procedure, not a physically structured compression procedure, and the

sparsity and size figures reported here should be read accordingly. Reconstructing the network with physically

removed channels and reporting its actual parameter count, memory footprint, and latency was not performed

in this pass and is noted as an item for future work in Section 8.2. Partial dynamic posttraining quantization

was applied to linear and LSTM submodules only because the dynamic-quantization path used does not support

1D convolutional layers; Conv1d weights remain in FP32. This is a toolchain limitation, not a design choice, and

the result is not equivalent to a full static INT8 model on a TFLite (Raspberry Pi 4) or TensorRT (Jetson Nano)

target. All conclusions drawn from the quantization results in Section 6.4 and Section 7 are limited to this

specific PyTorch/CPU partial-quantization implementation; Appendix A, Protocol C, specifies the full static

quantization procedure required for a target runtime. Knowledge distillation uses the standard Hinton-style

soft-target loss: a temperature-scaled Kullback–Leibler divergence between teacher and student log-softmax

outputs, combined with the student's own hard-label cross-entropy loss (Equation 9), with α = 0.5 and T = 4.0.

[12]

These values follow the defaults commonly used in the distillation literature; no hyperparameter sweep over

α or T was performed in this execution, and sensitivity to these choices is not characterized.

[12]

4.4 Experimental environment

All the experiments were run in a CPU-only sandbox (PyTorch, Python 3), with no GPU and no Raspberry Pi 4

or Jetson Nano hardware attached. Host-CPU latency figures reported in Section 6 were measured after 15

warm-up calls, using 100 timed single-sample inferences under a no-gradient context with wall-clock timing,

and are reported as a reproducibility and sanity check on the pipeline only; they are not representative of, and

must not be substituted for, Raspberry Pi 4 or Jetson Nano latency.

4.5 Decision threshold

All the configurations use a fixed decision threshold of 0.5 on the model's output probability to convert

predictions into positive/negative anomaly labels. This threshold was fixed prior to evaluation and was not

tuned on the validation or test sets; no threshold-optimization procedure is used in this manuscript. Because

this detail was not stated explicitly in the earlier version of this manuscript, the exact threshold value should be

confirmed by the authors against the underlying implementation before final publication (Section 10).

5. Real-time and energy measurement framework

This section defines the criteria against which the execution of the Appendix A protocol is evaluated. No

hardware numbers yet exist to substitute into these definitions; they are specified here so that the completed

study can be evaluated against a preregistered standard rather than a post hoc one.

5.1 Real-time criterion

For a 1 Hz vital sign monitoring scenario, the deadline is the sampling interval (1 s) minus any required

downstream action time. Compliance requires end-to-end latency (acquisition, preprocessing, inference, and

postprocessing, measured together on the target board) to fall below that deadline, not inference latency alone.

Real-time compliant ⇔ t_acquire + t_preprocess + t_inference + t_postprocess < T_deadline (1)

This manuscript's host-CPU inference latency (Table 4, 0.8–2.6 ms) is far below 1 s, but it is a generic-CPU figure

that excludes acquisition and preprocessing and is not Raspberry Pi 4 or Jetson Nano latency; it does not by

itself support a real-time claim on target hardware.

Table 4: Model size, parameter count, and host-CPU latency by configuration

Configuration

Parameters

Size (FP32/quantized)

Host-CPU latency, ms (mean ± SD)

M0 Baseline (CNN-LSTM)

62,562

252.0 KB

0.825 ± 0.054

M1 Pruning 10%

62,562

252.0 KB

0.846 ± 0.061

M1 Pruning 20%

62,562

252.0 KB

0.845 ± 0.067

M1 Pruning 30%

62,562

252.0 KB

0.822 ± 0.103

M1 Pruning 40%

62,562

252.0 KB

0.833 ± 0.057

M1 Pruning 50%

62,562

252.0 KB

0.803 ± 0.227

M2 Partial dynamic PTQ

62,562 (mixed precision)

85.1 KB

2.624 ± 0.148

M4 KD student (LightGRU)

1,662

9.1 KB

1.537 ± 0.192

M6 Pruning (30%) + KD

1,662

9.1 KB

1.541 ± 0.167

5.2 Energy calculation

E(J) = P(W) × t(s) (2)

E_inference (J/inference) = mean (P_active − P_idle) × mean(t_inference) (3)

Energy per inference and per monitoring window will be reported, once measured under Appendix A Protocol

B, as the mean ± SD across at least 100 repetitions and at least 3 independent power cycles, separating

inference-level, training-seed, and hardware-level variability.

5.3 Battery life and carbon

T_battery (h) = C_batt (Wh)/[(E_inference (J) × r_inference (1/h)/3600) + P_baseline (W)] (4)

M_carbon (g CO2) = E (kWh) × I_grid (g CO2/kWh) (5)

C_batt is the battery capacity, r_inference is the inference rate, P_baseline is the device's non-AI baseline draw,

and I_grid is a dated, cited grid carbon intensity figure for the deployment region. Appendix A, Protocol C,

specifies an appropriate Central Electricity Authority of India or state-grid emission-factor source for a

Chhattisgarh deployment context.

5.4 Compression and evaluation metrics

F1 = 2 × (Precision × Recall)/(Precision + Recall) (6)

Global sparsity = (zeroed parameters, whole model)/(total parameters, whole model) (7)

Compression ratio (params) = Params_baseline/Params_compressed (8)

L_KD = α × CE(y, σ(z_s)) + (1 − α) × T² × KL(σ(z_t/T) ‖ σ(z_s/T)) (9)

z_s and z_t are the student and teacher logarithms, respectively; σ is the softmax function; T is the distillation

temperature; and CE and KL denote the cross-entropy and Kullback–Leibler divergence, respectively.

6. Results

6.1 Configuration detection performance

All the values below were computed directly from the experimental evaluation outputs and independently

cross-checked from confusion-matrix quantities where verification was possible; no reported metric required

correction in this pass. The F1 score values from Table 5 for visual comparison across configurations are shown

in Fig. 3.

Table 5: Detection performance by configuration on the synthetic test set (360 windows)

Configuration

Accuracy

Precision

Recall

AUROC

M0 Baseline (CNN-LSTM)

0.978

0.977

0.934

0.955

0.993

M1 Pruning 10%

0.978

0.977

0.934

0.955

0.994

M1 Pruning 20%

0.969

0.976

0.901

0.937

0.991

M1 Pruning 30%

0.972

0.966

0.923

0.944

0.993

M1 Pruning 40%

0.983

0.978

0.956

0.967

0.998

M1 Pruning 50%

0.983

0.989

0.945

0.966

0.999

M2 Partial dynamic PTQ

0.978

0.977

0.934

0.955

0.993

M4 KD student (LightGRU)

0.956

0.912

0.986

M4 ablation: student, no KD

0.906

0.794

0.846

0.819

0.943

M6 Pruning (30%) + KD

0.944

0.949

0.824

0.882

0.967

Fig. 3: F1 score by configuration on the synthetic test set

6.2 Confusion matrices and specificity (derived)

To support the specificity and confusion matrix reported in the review, Table 6 reconstructs the per-

configuration true/false positive and negative counts and specificity from the accuracy, precision, and recall

values in Table 5 under the assumption of exactly 91 positive and 269 negative windows in the 360-window

test set (consistent with the 25.3% positive rate in Table 1). This reconstruction is internally consistent within

rounding for every configuration listed and is classified as a DERIVED quantity in Table 7, not a separately

measured quantity. A full precision–recall curve, and therefore an exact PR-AUC value, cannot be reconstructed

from these single-threshold summary metrics; the PR-AUC requires the underlying prediction scores, which are

not reproduced in this manuscript and are listed as outstanding items in Section 10.

Table 6: Confusion matrix counts and specificity by configuration

Configuration

Specificity

M0 Baseline (CNN-LSTM)

267

99.3%

M1 Pruning 10%

267

99.3%

M1 Pruning 20%

267

99.3%

M1 Pruning 30%

266

98.9%

M1 Pruning 40%

267

99.3%

M1 Pruning 50%

268

99.6%

M2 Partial dynamic PTQ

267

99.3%

M4 KD student (LightGRU)

261

97.0%

M4 ablation: student, no KD

249

92.6%

M6 Pruning (30%) + KD

265

98.5%

6.3 Model size, parameters, and host-CPU latency

The serialized model size values from Table 4 for visual comparison across configurations are shown in Fig. 4.

Table 7. Claim audit

Claim

Category

Baseline F1 = 0.955, AUROC = 0.993, 62,562 parameters

EXPERIMENTAL

LSTM holds 93.3% of parameters/93.7% of MACs

EXPERIMENTAL (recomputed)

Pruning whole-model sparsity 0.44–2.65% across 10–50% local

pruning

EXPERIMENTAL

KD student F1 = 0.912 vs. no-KD student F1 = 0.819 (0.093 gap)

EXPERIMENTAL

Confusion-matrix counts and specificity (Table 6)

DERIVED (reconstructed from Table 5 assuming 91/269

test-set composition)

Compression ratio (quantization) = 2.96x; MAC reduction (KD) =

40.8x

DERIVED

Global sparsity and compression-ratio formulas (Eq. 7, 8)

DERIVED (definitional)

Fixed 0.5 decision threshold, untuned on test set (Section 4.5)

Procedural detail, flagged for author confirmation

NEWS2 normal-range parameter table

Royal College of Physicians

[14]

MIMIC-III database description

Johnson et al.

[6]

DeepEdgeBench device/rate-dependent energy findings

Baller et al.

[7]

Structured-pruning parameter-concentration mechanism

Wen et al.

[5]

and He et al.

[8]

Knowledge-distillation loss formulation

Hinton et al.

[12]

Energy, power, battery-life, and carbon figures for this model

PROPOSED (Appendix A, Protocol B/C), not asserted

anywhere above

Real-time compliance on Raspberry Pi 4/Jetson Nano

PROPOSED, not asserted anywhere above

Detection performance on MIMIC-III/IV or WESAD data

PROPOSED (Appendix A, Protocol A), not asserted

anywhere above

PR-AUC

PROPOSED, requires archived prediction scores not

reproduced here

QAT (M3), M7, M8 results

PROPOSED, not executed, no result exists to classify

Fig. 4: Serialized model size by configuration

6.4 Pruning finding: Convolutional-layer pruning does not compress this model

Across pruning levels of 10–50% applied to conv1/conv2, the total parameter count and serialized size are

unchanged (62,562 parameters throughout, Table 4), and the host-CPU latency does not consistently decrease.

The measured whole-model sparsity (Equation 7) ranged from 0.44% at 10% local pruning to 2.65% at 50%

local pruning because the pruned convolutional layers hold only 6.5% of the total parameters (Table 2), and the

two-layer LSTM holds the remaining 93.3%. This is reported as a genuine, informative negative finding rather

than an experimental error: for CNN-LSTM or CNN-RNN hybrids generally, structured pruning must target the

recurrent layers, for example, through structured pruning methods designed specifically for recurrent weight

matrices or use low-rank factorization of the LSTM itself, to meaningfully shrink the model.

[5]

This

determination is empirical, is not carried out from the vision-pruning literature, and reflects the parameter-

concentration mechanism identified by structured-pruning surveys of convolutional architectures.

[8]

Reconstructing a physically pruned (rather than zeroed) network was not attempted in this pass; this

distinction between scarification and true structured compression is discussed further in Section 8.2. The

detection performance across pruning levels was stable and slightly improved at higher nominal sparsities (F1

increased from 0.955 at 10% to 0.967 at 40%), which is consistent with pruning acting as a mild regularizer on

this architecture. Given the single-seed protocol used throughout this study (Section 8.1), this rise cannot be

distinguished from run-to-run noise on the strength of this result alone; a regularization interpretation is

plausible but unconfirmed, and Appendix A's multiseed repetition is required before treating it as a real effect.

This stability is not a compression benefit in any case because the model was not actually compressed in a way

that reduces deployed size or (unmeasured) energy; accuracy under pruning and size under pruning are

distinct effects that must not be conflated.

6.5 Quantization (partial dynamic PTQ)

Partial dynamic INT8 quantization of the linear and LSTM submodules reduced the serialized size from 252.0

KB to 85.1 KB, a 66.2% reduction (Equation 8: compression ratio = 2.96x), with the detection metrics essentially

unchanged (F1 = 0.955 versus baseline 0.955; Table 5). The host-CPU latency for the quantized path was higher,

not lower, than that of FP32 (2.62 ms versus 0.82 ms), an expected artifact of dynamic quantization per-call

quantize/dequantize overhead on a general-purpose CPU lacking INT8-optimized kernels for this operator set

in this specific PyTorch/CPU implementation, and no evidence that quantization is harmful in general. Static

INT8 on a TFLite or TensorRT target (Appendix A) is expected to behave differently in terms of both size and

latency and must be measured, not inferred, from this CPU-sandbox result.

6.6 Knowledge distillation

Knowledge distillation to the 1,662-parameter GRU student (9.1 KB; 90,040 MACs, a 40.8x reduction relative to

the teacher's 3,673,088 MACs) reached F1 = 0.912, AUROC = 0.986. With matched ablation, the identical student

architecture trained with hard labels only (α = 0 in Equation 9) reached F1 = 0.819 and AUROC = 0.943. The

0.093-point F1 difference is attributable to distillation, with all other factors held fixed, from an ablation run

explicitly seeking to measure this quantity. As with all the point estimates in this manuscript, this difference

reflects a single training seed (42) and should not be interpreted as a statistically validated effect size; Appendix

A specifies multiseed repetition as part of the extension protocol.

6.7 Combined pruning and knowledge distillation

Distilling from a 30%-pruned teacher (M6) produced F1 = 0.882, which is below the plain-teacher distillation

result (F1 = 0.912; Section 6.6). This is reported as follows: combining pruning and distillation did not

outperform distillation alone in this single run. With only one run of this combination, this should be read as an

observed result rather than a statistically confirmed interaction effect.

6.8 Not executed or not measured

Table 8 shows the items specified in the underlying study plan that were not executed or measured in this pass.

Table 8: Planned protocol items omitted or unmeasured in the current pass

Item

Status

M3 (QAT)

Not executed. Requires a locked target runtime (TFLite for Raspberry Pi

4/TensorRT for Jetson Nano) Appendix A, Protocol C

M7 (PTQ+KD), M8 (Pruning+QAT+KD)

Not executed this pass

Raspberry Pi 4/Jetson Nano latency, power,

energy per inference

Not measured. No physical hardware available Appendix A, Protocol B

Battery life

Not calculable. Depends on measured inference energy

Carbon emissions

Not calculable. Depends on measured energy and the cited grid carbon-intensity

figure

Statistical significance testing across seeds

Not run this pass. Single seed (42); Appendix A specifies at least 3-seed repetition

MIMIC-III/IV or WESAD-based results

Not available. No credentialed data access in this environment

PR-AUC

Not computable from reported summary metrics. Requires archived prediction

scores

7. Discussion

• RQ1: Which compression method works best for detection quality on this synthetic benchmark?

Knowledge distillation gave the clearest, ablation-verified benefit, an approximately 41x reduction in

the MAC, while most of the accuracy gap of the compact model was recovered relative to that of the

undistilled student (Section 6.6).

• RQ2: Does reducing local (per-layer) sparsity produce meaningful whole-model compression here? No.

Local channel sparsity up to 50% in the convolutional layers produced only 0.44–2.65% whole-model

sparsity because the LSTM, not the convolutional front end, dominated both the parameters and the

MACs (Section 6.4, Table 2). Local and global sparsity must be reported separately for architectures

with this kind of parameter concentration.

• RQ3: Does quantization provide meaningful computational benefit here? In part, with an important

caveat specific to this implementation, the model’s storage footprint was genuinely reduced by 66.2%

under partial dynamic quantization, but host-CPU latency increased owing to the dynamic quantization

overhead of this specific PyTorch/CPU configuration (Section 6.5). This conclusion is scoped to that

implementation and should not be generalized to static INT8 quantization on a target runtime, which

Appendix A specifies as a separate, unexecuted evaluation.

• RQ4: Does knowledge distillation allow a much smaller student to retain useful detection performance?

Yes, the matched ablation in Section 6.6 shows a clear, controlled benefit from distillation; as a single-

seed result, this should be read as a strong preliminary finding rather than a statistically confirmed one.

•

RQ5: Which method provides the best energy reduction? Unknown. No energy was measured (Section

6.8). Parameter or MAC reduction is not a substitute for measured energy: DeepEdgeBench's energy

findings, which depend on both the device and inference rate, show that a FLOP count alone would not

predict measured energy behavior, so this manuscript deliberately does not extrapolate from MACs to

energy.

[7]

• RQ6: What was the accuracy cost of compression? For distillation, the real cost was substantially,

although not completely, recovered (Section 6.6). For pruning as implemented, negligible accuracy cost

but also negligible compression (Section 6.4); these two effects should not be read as offsetting one

another, since only one of them occurred at meaningful scale.

• RQ7: Did the combined configuration (pruning + distillation) outperform distillation alone? No: In the

one combined configuration run (M6), the performance was below that of plain distillation (Section

6.7). With only one run of this combination, this is reported as an observed result, not a confirmed

interaction effect.

• RQ8: Can the compressed model meet the defined real-time constraint? This is not yet demonstrable

because the definition in Section 5.1 requires end-to-end, on-device latency, which was not measured.

None of the results above support a deployment recommendation on physical hardware: the

compression pipeline, labeling logic, and ablation design are validated on synthetic data, while the

clinical and hardware claims remain unestablished pending the Appendix A protocol.

8. Threats to validity and limitations

8.1 Threats to validity

• Construct validity: NEWS2-band threshold crossing is a physiological-abnormality proxy, not a clinical

diagnosis or a NEWS2 aggregate score; no claim of diagnostic capability is made anywhere in this

manuscript, and the high F1/AUROC values reported reflect detectability of a deterministic, rule-based

synthetic label rather than evidence of clinical effectiveness.

• Internal validity: A single random seed (42) was used throughout; the point estimates in Tables 4-6

should not be read as statistically validated differences between configurations.

• External validity: The synthetic generator’s AR(1) drift and its anomaly structure of one episode per

window do not reflect real ICU comorbidity, missingness, or multiple-episode dynamics; generalization

to MIMIC-III/IV or WESAD data is untested.

• Ecological validity of latency figures: all latency numbers are generic-CPU sandbox measurements and

are not representative of Raspberry Pi 4 or Jetson Nano behavior.

8.2 Limitations

• Synthetic-data proof-of-concept only: no MIMIC-III/IV or WESAD access was available in this

environment, and the results should not be read as clinical evidence.

• No physical edge hardware was available: no Raspberry Pi 4, Jetson Nano, or INA219 sensor was

available.

• Sparsification, not structured compression: The pruning API used zero weights without physically

shrinking tensors; reconstructing the network with physically removed channels and reporting its

actual parameter count, memory footprint, and latency was not performed and is a priority for future

work.

• Recurrent-layer pruning is not attempted: given that the LSTM holds 93.3% of the parameters (Section

4.2), meaningful compression of this architecture likely requires pruning strategies designed for

recurrent weight matrices, reduced hidden dimensions, or low-rank factorization, none of which was

attempted in this execution.

[5]

• Partial quantization scope: Only linear/LSTM submodules were quantized (Conv1d unsupported by

this quantization path), which is not equivalent to a full static INT8 conversion on a real deployment

runtime.

• QAT (M3), M7, and M8 have not been executed; no conclusions about the fully combined pipeline are

implied.

• Generalizability beyond this specific synthetic generator and this specific CNN-LSTM/GRU architecture

pair is untested.

9. Claim audit

Every major claim in this manuscript is classified below into one of four categories: EXPERIMENTAL (produced

and independently verified), DERIVED (calculated from experimental values), LITERATURE (supported by a

cited external source), or PROPOSED (specified but not yet executed), as illustrated in Table 7. No PROPOSED

claim is presented elsewhere in this manuscript as an EXPERIMENTAL finding.

10. Reproducibility and data/code availability

The experimental pipeline comprises three components: dataset generation; model/architecture definitions;

and a combined training, compression, and evaluation script. The software environment used Python 3 with

PyTorch (CPU build) and scikit-learn. The hyperparameters, random seed (42), 0.5 decision threshold (Section

4.5), and early stopping criteria are fixed constants set prior to running the experiments and not tuned post

hoc. Dataset generation, patient-level split, and normalization fitting are deterministic given the seed.

All values reported in Tables 1–6 were produced directly by the experimental pipeline's evaluation outputs,

with parameter and MAC counts additionally recomputed independently from the architecture definitions as a

cross-check; no reported metric required correction in this pass. No proprietary data or code dependency

exists. The complete code, configuration, and result records should be made available through a public

repository at the time of resubmission; the repository URL is not yet included in this manuscript and is listed

as an outstanding item in Section 10's companion list below. The exact AR (1) coefficient and noise standard

deviation used by the synthetic-data generator should also be published alongside the repository to allow exact

reproduction of Table 1. The only remaining prerequisite for reproducing the full target study described in

Appendix A is credentialed MIMIC/WESAD data access and the physical edge hardware specified there.

11. Conclusion

This execution establishes a validated, reproducible compression pipeline and a clinically grounded, NEWS2-

based labeling methodology for vital-sign anomaly detection on synthetic data, and its detection metrics should

be read as characterizing that synthetic-data proof-of-concept rather than clinical performance. This finding

clearly confirms that knowledge distillation meaningfully recovers accuracy in a compact GRU student,

alongside one clear negative finding: naive convolutional pruning does not meaningfully compress an LSTM-

dominated architecture, because the LSTM holds 93.3% of the parameters and achieved whole-model sparsity

remaining below 2.65% across the pruning sweep. This manuscript does not, and does not claim to, establish

the energy-aware, real-time, hardware-validated, MIMIC-based results implied by a fully realized version of this

study. Appendix A specifies exactly what is required to complete that study (PhysioNet credentialing, a

Raspberry Pi 4, a Jetson Nano, and an INA219 power sensor) without redesigning any of the software developed

here.

Acknowledgment

Not applicable

CRediT Author Contribution Statement

Pushpak Basak: Conceptualization, Methodology, Software, Writing – original draft, Writing – review & editing,

Supervision. Sureeti Sahu: Data curation, Investigation, Writing – review & editing. All the authors have read

and approved the final version of the manuscript for publication and agree to be accountable for all aspects of

the work, ensuring that questions related to the accuracy or integrity of any part of the work are appropriately

investigated and resolved.

Funding Declaration

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-

profit sectors.

Data Availability Statement

The datasets generated and/or analyzed during the current study that support the findings are available from

the corresponding author upon reasonable request.

Conflict of Interest

There are no conflicts of interest.

Artificial Intelligence (AI) Use Disclosure

The authors declare that artificial intelligence (AI)-assisted tools were used only for language refinement,

grammar improvement, and manuscript structuring purposes during the preparation of this work. All technical

content, experimental implementation, results, and interpretations were independently developed and verified

by the authors.

Supporting Information

Appendix A: Protocol for Clinical Data and Edge Hardware Extension

This appendix contains, in one place, the extension protocol mentioned throughout the main text. It describes

what needs to be done to finish the target study called for in the manuscript's Introduction; none of the

following steps have been done, and no results from them claimed anywhere in the main manuscript.

Protocol A: Real clinical data substitution

The synthetic-data generator is replaced with a MIMIC-III/IV- or WESAD-derived extraction, preserving the

same (features, labels, patient ID) interface as that consumed by the downstream training and evaluation code

so that no changes to the compression pipeline itself are needed. This requires (i) PhysioNet credentials and

data-use agreement completion for MIMIC-III/IV and/or WESAD; (ii) a definition of an anomaly label consistent

with or compared to the NEWS2-band threshold rule used in the main manuscript; and (iii)

training/validation/testing data split at the patient level with no patient overlap between splits, analogous to

the way the synthetic data were held out; and (iv) explicit handling of irregular sampling, missing values, and

multiple episode events, features of which are absent in the synthetic generator.

Protocol B: Physical edge-hardware measurement

The inference latency (and power draw, if applicable) was measured on a Raspberry Pi 4 and an NVIDIA Jetson

Nano (or other devices), and the device and inference rate were varied according to a protocol similar to that

used by DeepEdgeBench (rather than using a fixed device ranking).

[7]

This involves (i) deployment of the

trained model(s) to each target board; (ii) an INA219 (or equivalent) power sensor to measure the power draw

(active and idle); (iii) at least 100 timed inferences and at least 3 independent power cycles per configuration

(mean ± SD); and (iv) end-to-end latency (acquisition, preprocessing, inference, postprocessing) measurement

to evaluate the end-to-end criterion defined in the main manuscript (Section 5.1), rather than the inference-

only latency.

Protocol C: Target-runtime quantization, battery life, and carbon accounting

We execute quantization-aware training (M3) and a full static INT8 inference pipeline on a constrained

deployment runtime (TFLite for the Raspberry Pi 4, TensorRT for the Jetson Nano) rather than the CPU-only

partial dynamic quantization reported in the main manuscript (Section 6.5) and rerun the PTQ+KD (M7) and

pruning+QAT+KD (M8) configurations once M3 has converged. Using the measured energy values from

Protocol B, calculate battery life and carbon emissions using the formulas defined in the main manuscript

(Section 5.3), citing a dated Central Electricity Authority of India or state-grid emission-factor source suitable

for a Chhattisgarh deployment context.

Statistical repetition

Repeat the full ablation (main manuscript, Table 3) across at least three random seeds once Protocols A–C are

executed to facilitate statistical comparisons between configurations; single-seed point estimates, as reported

in the main manuscript, should not be taken as valid statistically significant differences.

Appendix B: Outstanding Items for Author Verification

The following items are cited explicitly in the main manuscript but cannot be determined without the authors'

own implementation records. These have no effect on any of the results reported in the main manuscript.

• Configuration identifier M5 (Table 3 of the main manuscript): No configuration named M5 appears in

the study plan or execution record. It is unclear if M5 was deliberately skipped in the original

configuration numbering, reserved for a configuration that was subsequently merged into M6, or

otherwise omitted by error.

• Synthetic-generator parameters (Section 4.1 of the main manuscript): the actual value of the AR (1)

coefficient and noise standard deviation used by the generator must be published alongside the code

repository to reproduce Table 1 exactly.

• Decision threshold (Section 4.5 of the main manuscript): the manuscript says a fixed 0.5 threshold was

used and not tuned on the test set; this must be verified against the underlying implementation.

• PR-AUC: This metric cannot be reconstructed from the single-threshold precision/recall values

reported in the main manuscript's Table 5; it should be calculated and reported from the archived

prediction scores.

• Public repository: the code, configuration, and result records should be archived in a public repository;

the repository URL should be included at the time of resubmission

Appendix C: Figure Description (System and Model Architecture)

The system and model architecture are presented in Fig. 1 in the main manuscript, Section 4.2. For reference,

here is the text version of the content: the baseline/teacher pipeline takes as input a window of 5 channels × 60

timesteps (heart rate, respiratory rate, Spo2, systolic blood pressure, temperature) through two 1D

convolutional blocks (32 channels, kernel sizes 5 and 3, with batch normalization and ReLU after each), a two-

layer LSTM (64 hidden units per layer, dropout 0.2), a dropout layer (0.3), and a fully connected classification

head, predicting an anomaly probability against a fixed 0.5 decision threshold. Three compression paths

originate from this pipeline: M1 (structured channel pruning of the convolutional layers), M2 (partial, dynamic

posttraining INT8 quantization of linear and LSTM layers only), and M4 (knowledge distillation to a separate 1-

layer, 20-hidden-unit GRU student, 'LightGRU,' 1662 parameters, 90040 MACs). M6 is the combination of M1

and M4, with distillation from a 30%-pruned teacher.

References

[1]

T. Liang, J. Glossner, L. Wang, S. Shi, X. Zhang, Pruning and quantization for deep neural network

acceleration: A survey, Neurocomputing, 2021, 461, 370–403, doi: 10.1016/j.neucom.2021.07.045.

[2]

S. Muralitharan, W. Nelson, S. Di, M. McGillion, P. J. Devereaux, N. G. Barr, J. Petch, Machine learning–

based early warning systems for clinical deterioration: Systematic scoping review, Journal of Medical

Internet Research, 2021, 23, e25187, doi: 10.2196/25187.

[3]

R. E. Ko, Z. Kim, B. Jeon, M. Ji, C. R. Chung, G. Y. Suh, M. J. Chung, B. H. Cho, Deep learning-based early

warning score for predicting clinical deterioration in general ward cancer patients, Cancers, 2023, 15,

5145, doi: 10.3390/cancers15215145.

[4]

M. S. Diab, E. Rodriguez-Villegas, Embedded machine learning using microcontrollers in wearable and

ambulatory systems for health and care applications: A review, IEEE Access, 2022, 10, 98450–98474,

doi: 10.1109/ACCESS.2022.3206782.

[5]

L. Wen, X. Zhang, H. Bai, Z. Xu, Structured pruning of recurrent neural networks through neuron

selection, Neural Networks, 2020, 123, 134–141, doi: 10.1016/j.neunet.2019.11.018.

[6]

A. E. W. Johnson, T. J. Pollard, L. Shen, L. H. Lehman, M. Feng, M. Ghassemi, B. Moody, P. Szolovits, L. A.

Celi, R. G. Mark, MIMIC-III, a freely accessible critical care database, Scientific data, 2016, 3, 160035,

2016, doi: 10.1038/sdata.2016.35.

[7]

S. P. Baller, A. Jindal, M. Chadha, M. Gerndt, DeepEdgeBench: Benchmarking deep neural networks on

edge devices, In 2021 IEEE international conference on cloud engineering (IC2E), IEEE, 2021, 20–30,

doi: 10.1109/IC2E52221.2021.00016.

[8]

Y. He, L. Xiao, Structured pruning for deep convolutional neural networks: A survey, IEEE transactions

on pattern analysis and machine intelligence, 2023, 46, 2900–2919, doi:

10.1109/TPAMI.2023.3334614.

[9]

T. Suwannaphong, F. Jovan, I. Craddock, R. McConville, Optimising TinyML with quantization and

distillation of transformer and Mamba models for indoor localisation on edge devices, Scientific

Reports, 2025, 15, 10081, 2025, doi: 10.1038/s41598-025-94205-9.

[10]

X. An, S. Shi, Q. Wang, Y. Yu, Q. Liu, Research on a lightweight arrhythmia classification model based on

knowledge distillation for wearable single-lead ECG monitoring systems, Sensors, 2024, 24, 7896, doi:

10.3390/s24247896.

[11]

A. John, B. Cardiff, D. John, A 1D-CNN based deep learning technique for sleep apnea detection in IoT

sensors, In Proc. 2021 IEEE Int. Symp. on Circuits and Systems (ISCAS), IEEE, 2021, 1–5, doi:

10.1109/ISCAS51556.2021.9401300.

[12]

G. Hinton, O. Vinyals, J. Dean, Distilling the knowledge in a neural network, 2015, arXiv:1503.02531,

doi: 10.48550/arXiv.1503.02531.

[13]

R. Schwartz, J. Dodge, N. A. Smith, O. Etzioni, Green AI, Communications of the ACM, 2020, 63, 54–63,

doi: 10.1145/3381831.

[14]

Royal College of Physicians, National Early Warning Score (NEWS) 2: Standardising the Assessment of

Acute-Illness Severity in the NHS, Updated Report. London, U.K.: RCP, 2017,

https://www.rcp.ac.uk/media/umzn4ntq/news2_additional-guidance-002-_0.pdf, Accessed 07

August 2026.

Publisher Note: The views, statements, and data in all publications solely belong to the authors and

contributors. GR Scholastic is not responsible for any injury resulting from the ideas, methods, or products

mentioned. GR Scholastic remains neutral regarding jurisdictional claims in published maps and institutional

affiliations.

Open Access

This article is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License, which

permits the non-commercial use, sharing, adaptation, distribution and reproduction in any medium or format,

as long as appropriate credit to the original author(s) and the source is given by providing a link to the Creative

Commons License and changes need to be indicated if there are any. The images or other third-party material

in this article are included in the article's Creative Commons License, unless indicated otherwise in a credit line

to the material. If material is not included in the article's Creative Commons License and your intended use is

not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly

from the copyright holder. To view a copy of this License, visit: https://creativecommons.org/licenses/by-

nc/4.0/