AI on the Pulse: Adaptive Real-Time Analysis

• AI on the Pulse is a paradigm that integrates real-time signal processing with adaptive learning to continuously monitor and respond to evolving data streams.
• It employs hybrid models, sensor fusion, and dynamic feedback loops to enhance diagnostics, anomaly detection, and optimization across diverse domains.
• The approach spans applications from biomedical auscultation and wearable monitoring to satellite imaging and ultrafast physics, emphasizing continuous adaptation and personalized baselines.

Searching arXiv for the cited works to ground the article and confirm metadata. “AI on the Pulse” denotes a family of AI systems that operate on pulse-like, temporally evolving, or real-time signals and pair inference with responsiveness to changing context. Across medicine, humanitarian mapping, astrophysics, accelerator physics, ultrasound tomography, customer-experience systems, and materials science, the phrase consistently refers to AI that remains tightly coupled to live signals, changing environments, or sequential human decision loops rather than relying only on static offline prediction (Ghouse et al., 18 May 2025). In biomedical settings, this includes direct analysis of auscultation sounds, wearable physiological streams, echocardiographic motion, photoplethysmography, and behavioral attention signals; in other domains, it includes adaptive satellite-image analysis, pulse-shape reconstruction, pulse-parameter optimization, and continuously updated deployment evaluation (Gabrielli et al., 5 Aug 2025). A common thread is the replacement of one-shot, closed-world inference with pipelines that sense, model, adapt, and often preserve a digitally reviewable trace of the underlying signal or decision state (Logar et al., 2020).

1. Cardiopulmonary sensing at the point of care

One concrete instantiation of AI on the pulse is the use of AI-enhanced auscultation for cardiopulmonary diagnosis. In “AI- Enhanced Stethoscope in Remote Diagnostics for Cardiopulmonary Diseases” (Ghouse et al., 18 May 2025), a traditional acoustic stethoscope is augmented with a microphone physically attached to the chest piece and connected via a 3.5 mm audio jack to a laptop, tablet, or smartphone. The system records heart and lung sounds, performs data preprocessing, extracts MFCCs and Mel spectrograms, and classifies recordings with a hybrid CNN+GRU model integrated into a Streamlit web app (Ghouse et al., 18 May 2025).

The clinical motivation is explicit. Cardiovascular and pulmonary diseases are described as leading causes of death worldwide, while traditional auscultation remains limited by subjectivity, skill dependence, lack of persistent digital record, and restricted access to trained clinicians in under-resourced regions (Ghouse et al., 18 May 2025). The proposed system addresses these constraints by producing standardized digital audio, automated classification, and a replayable longitudinal record that can be emailed to a clinician for remote review (Ghouse et al., 18 May 2025). This supports triage, telemedicine, and follow-up rather than replacing clinical oversight.

Technically, the signal-processing pipeline is tightly specified. A 2nd-order Butterworth band-pass filter with cut-off frequencies at 25 Hz and 400 Hz is applied, the signal is resampled to 1000 Hz, truncated to the first 2500 samples, and normalized to the range $(-1,1)$ (Ghouse et al., 18 May 2025). For underrepresented lung classes in the ICBHI 2017 dataset, augmentation is performed using artificial noise, temporal shifting, and pitch and speed changes (Ghouse et al., 18 May 2025). The main features are MFCCs, with the standard FFT, Mel filterbank, logarithmic compression, and DCT pipeline, while the network input accepts MFCC features with shape $(1,52)$ (Ghouse et al., 18 May 2025).

The hybrid model combines two convolutional blocks and five GRU layers. The first convolution uses kernel size 11 with 256 filters, followed by max pooling, batch normalization, and ReLU; the second uses 512 filters with the same pattern (Ghouse et al., 18 May 2025). The GRU stack uses unit sizes 32, 64, and 128 with tanh activation, after which flattening and dense layers with Leaky ReLU lead to an 11-unit softmax output for 6 lung and 5 heart classes (Ghouse et al., 18 May 2025). Labels are one-hot encoded, the optimizer is Adam with learning rate 0.0002, the loss is categorical cross-entropy, the batch size is 8, and training runs up to 150 epochs with Early Stopping and checkpointing (Ghouse et al., 18 May 2025).

The reported performance is 94% average accuracy for both lung and heart classification with the CNN+GRU hybrid, outperforming CNN-only and GRU-only baselines (Ghouse et al., 18 May 2025). On lung sounds, the hybrid reaches 94% average accuracy, with classwise values including URTI precision 0.90, recall 0.88, F1 0.89, accuracy 0.93, and pneumonia precision 0.65, recall 0.60, F1 0.62, accuracy 0.75 (Ghouse et al., 18 May 2025). On heart sounds, average accuracy is likewise 94%, with approximately 94% per-class accuracy across AS, MS, MR, MVP, and normal classes (Ghouse et al., 18 May 2025). This suggests that the phrase “AI on the pulse,” in one of its most literal senses, can refer to AI operating directly on chest-wall biosignals for low-cost, real-time remote diagnostics.

A related but distinct cardiology example appears in the “Acoustic Index” study (Begiashvili et al., 17 Jul 2025), which defines a continuous score from 0 to 1 for cardiac disease risk derived from echocardiographic sequences and clinical metadata. There the pulse is not acoustic in the auscultatory sense but mechanical and dynamical: the beating heart is modeled via Extended Dynamic Mode Decomposition and Koopman operator theory, with spatiotemporal modes weighted by attention and fused with clinical data on a latent manifold (Begiashvili et al., 17 Jul 2025). In a prospective cohort of 736 patients, the Acoustic Index achieved an AUC of 0.89 on an independent test set, with sensitivity and specificity exceeding 0.8 near a threshold of 0.45 (Begiashvili et al., 17 Jul 2025). This broadens the phrase from audio pulse signals to video-derived cardiac dynamics.

2. Continuous physiological monitoring and individualized anomaly detection

A second major use of AI on the pulse is continuous home monitoring via wearables and ambient sensors. “AI on the Pulse: Real-Time Health Anomaly Detection with Wearable and Ambient Intelligence” (Gabrielli et al., 5 Aug 2025) defines an end-to-end anomaly detection system built around UniTS, a universal time-series foundation model, to learn each patient’s normal physiological and behavioral patterns and then flag deviations in real time. The system uses smartwatch-like signals such as HR, HRV, CVRR, respiration rate, SpO$_2$, blood pressure, steps, and sleep phases, together with ambient variables including temperature, humidity, CO$_2$, air pressure, luminosity, TVOC, and room location, all aligned at 1-minute resolution (Gabrielli et al., 5 Aug 2025).

The paper explicitly motivates anomaly detection over classification. Dense labeling of “stress,” “health event,” or “no event” at every time point is described as impractical in real-world home deployment, especially because anomalies are patient-specific, rare, and subject to baseline drift (Gabrielli et al., 5 Aug 2025). Instead, the model is trained on typical behavior, reconstructs masked future windows conditioned on recent history and context, and uses the reconstruction error as an anomaly score (Gabrielli et al., 5 Aug 2025). Missing data are handled by nearest-window interpolation when no more than 5 samples are missing within a 16-sample window, after comparison against other interpolation methods using MASE and DTW (Gabrielli et al., 5 Aug 2025).

UniTS is adapted to contextual anomaly detection by framing the problem as masked future imputation. In the @HOME setting, the sliding window length is $K=16$, with the first 8 samples unmasked and the last 8 masked and reconstructed (Gabrielli et al., 5 Aug 2025). The architecture uses sample tokens, context tokens, prompt tokens, dual self-attention across time and features, dynamic feed-forward networks, gating modules, and a Tower module that reconstructs $\hat{\mathbf{X}}_F$ from future embeddings (Gabrielli et al., 5 Aug 2025). Thresholding in deployment is patient-specific and based on Peaks-Over-Threshold from extreme value theory (Gabrielli et al., 5 Aug 2025).

The comparative evaluation spans 12 anomaly detection baselines on DREAMER, HCI, WESAD (ECG), and WESAD (BVP). UniTS attains an average F1 of $0.821 \pm 0.049$ versus the best baseline, DAGMM, at $0.673 \pm 0.066$, corresponding to an approximately 22% relative improvement (Gabrielli et al., 5 Aug 2025). Per-dataset results include F1 0.828 on DREAMER, 0.864 on HCI, 0.793 on WESAD (ECG), and 0.800 on WESAD (BVP), with FPR approximately 0.031 on both WESAD variants (Gabrielli et al., 5 Aug 2025). The small ECG-to-BVP degradation is presented as evidence that consumer wearables can support meaningful anomaly detection.

The real-world deployment is equally central. In @HOME, 6 elderly patients with early-stage neurological conditions were monitored continuously at 1-minute resolution, yielding 32 detected anomalies over the first 3 months: 10 hyper/hypotension, 9 abnormal HRV, 10 stress, and 3 sleep-quality anomalies, of which 2 were sensor errors (Gabrielli et al., 5 Aug 2025). An experienced senior geriatrician confirmed 30 of 32 anomalies as true positives, for a 93.75% confirmation rate (Gabrielli et al., 5 Aug 2025). The same system integrates LLM-generated natural-language explanations for clinicians, translating anomaly traces into short medically worded summaries that remain subordinate to clinician review (Gabrielli et al., 5 Aug 2025). Here “AI on the pulse” denotes minute-scale, personalized, unsupervised surveillance of physiologic and environmental patterns in the home.

A more behaviorally grounded counterpart appears in “Pulse Focus” (Debele et al., 2 Jun 2026), where the Focus Performance Score is derived from a mobile Stroop task and validated behaviorally and neurally as a 0–100 attentional control score. Behavioral validation on 466 adults and 111,133 trials showed that FPS captures Stroop interference and has test–retest ICCs of 0.831, 0.921, and 0.928 across sessions, while neural validation on the DMCC55B dataset linked mean incongruent reaction time, a primary FPS component, to ACC activation $(\rho=-0.327, p=0.015)$ (Debele et al., 2 Jun 2026). This suggests a broader interpretation of “pulse” as a measurable cognitive pulse or behavioral readout for attention-aware AI, rather than only a cardiovascular signal.

3. Signal fusion and pulse-shape inference in biomedical and physical systems

The phrase also appears in settings where the central technical problem is fusion of multiple pulse-like signals or reconstruction of pulse shape under distortion. In wrist-based heart-rate monitoring, PULSE is the name of a deep model using temporal convolutions and multi-head cross-attention to fuse PPG and accelerometer signals (Kasnesis et al., 2022). Inputs are standardized 8-second windows, downsampled to 32 Hz, with shape $4 \times 256$ for PPG-DaLiA and $(1,52)$0 for the IEEE datasets, followed by per-channel z-score normalization (Kasnesis et al., 2022). The architecture applies three 1D convolutional blocks with output channels 32, 48, and 64, uses the PPG embeddings as queries and the 3-axial accelerometer embeddings as keys and values in a 4-head cross-attention module, and then regresses a scalar HR estimate through two dense layers (Kasnesis et al., 2022).

The results show that sensor fusion is not automatically beneficial unless it is performed in a way that reflects cross-modal dependencies. In the ablation reported in the provided details, PPG-only with MHSA achieved MAE 4.52 BPM, PPG+accelerometer with MHSA worsened to 5.11 BPM, while the cross-attentional fusion mechanism is presented as the effective way to exploit motion signals (Kasnesis et al., 2022). The paper states that on PPG-DaLiA, PULSE reduces mean absolute error by 7.56% relative to previous work (Kasnesis et al., 2022). The attention maps are further used as an explainability device, revealing which motion channels and times are influencing the HR estimate.

In resuscitation, “Machine Learning and Feature Engineering for Predicting Pulse Status during Chest Compressions” (Sashidhar et al., 2020) addresses a different pulse-related ambiguity: whether organized ECG activity during CPR corresponds to a spontaneous pulse. The dataset contains 383 OHCA patients, with 540 pulse checks in the training set and 372 in the test set, of which 38% overall had a spontaneous pulse (Sashidhar et al., 2020). The method resamples ECG to 250 Hz, applies a 4th-order Butterworth bandpass filter from 1 to 40 Hz, computes a bump-wavelet scalogram, reduces it via PCA to three principal modes, and classifies pulse status using linear discriminant analysis (Sashidhar et al., 2020). The AUC on test data is 0.84 during CPR and 0.89 without CPR (Sashidhar et al., 2020). This is “AI on the pulse” in a clinically operational sense: inferring perfusion without pausing chest compressions.

The notion extends well beyond biomedicine. In free-electron laser diagnostics, VPuRD reconstructs single-shot FEL pulse power using a virtual diagnostic based on an MLP with 23 scalar machine parameters as input and a 567-point temporal power profile as output (Korten et al., 2024). The training set contains 2,826 experimental lasing-off shots, split into 2,261 training, 283 validation, and 282 test shots (Korten et al., 2024). By combining the predicted lasing-off electron temporal power profile with the measured lasing-on profile, the method reconstructs photon power on a single-shot basis, outperforming both a mean-profile baseline and neighboring-shot substitution (Korten et al., 2024). Here the pulse is an ultrafast radiation pulse, and the AI acts as a non-invasive software instrument.

Likewise, in ultrafast optics, AI-assisted spectral phase modulation is used to optimize femtosecond pulse amplification without any training dataset or explicit model of the amplifier (Krakowski et al., 2023). A Grey Wolf Optimizer adjusts the spectral phase on an SLM and evaluates pulse quality using the fitness $(1,52)$1 derived from autocorrelation traces (Krakowski et al., 2023). The best configuration reduces FWHM from 46 fs to 36 fs and lowers side-pulse intensity by a factor of 4.6 (Krakowski et al., 2023). This is a literal example of AI operating on a physical pulse by directly sculpting its spectral phase.

4. Human–AI feedback loops and dynamic world modeling

Another major strand of AI on the pulse is not about physiological or physical pulses at all, but about staying on the pulse of evolving environments through continuous adaptation. “PulseSatellite” (Logar et al., 2020) is exemplary in this regard. It is a web-based humanitarian satellite image analysis platform built by UN Global Pulse and UNOSAT that integrates neural models with human correction and on-the-fly retraining (Logar et al., 2020). For settlement mapping it uses Mask R-CNN trained on 12 refugee settlements tiled into 300×300 pixel images, and for flood mapping it uses U-Net trained on flood events in Bangladesh and Somalia (Logar et al., 2020).

The central mechanism is a human–AI feedback loop. Analysts run a model on new imagery, correct predictions on selected tiles, trigger adaptation, and then re-run the adapted model in the same operational session (Logar et al., 2020). In shelter mapping, this improves camp completion rates from 77.3% before adaptation to 94.7% after adaptation, with a final user accuracy of 94.4% (Logar et al., 2020). In flood mapping, the U-Net achieves overall accuracy greater than 90%, with maps available in minutes rather than hours (Logar et al., 2020). This version of being “on the pulse” is about keeping AI coupled to the unfolding specifics of each humanitarian crisis.

A more abstract deployment-centric counterpart is “AgentPulse” (Gao et al., 27 Apr 2026), which argues that static benchmarks capture what agents can do at a point in time but not how they are adopted, maintained, or experienced in the wild. AgentPulse defines a four-factor composite over 50 agents and 18 real-time signals: Benchmark Performance, Adoption Signals, Community Sentiment, and Ecosystem Health (Gao et al., 27 Apr 2026). The factors are complementary, with the maximum pairwise correlation being Adoption–Ecosystem at $(1,52)$2, while Benchmark–Adoption is only $(1,52)$3 (Gao et al., 27 Apr 2026). A circularity-controlled test on 35 agents shows that a Benchmark+Sentiment sub-composite predicts external adoption proxies not used in the sub-composite, including GitHub stars $(1,52)$4 and Stack Overflow question volume $(1,52)$5 (Gao et al., 27 Apr 2026). This suggests that “AI on the pulse” can also denote continuous measurement of deployment reality.

In customer-experience systems, PulseCX breaks what it calls the Closed-World Constraint by decoupling knowledge acquisition from knowledge consumption (Agarwal et al., 19 Jun 2026). An asynchronous Social Search Agent linearizes external signals into Context Objects, which are stored in a Decay-Aware Temporal Knowledge Graph governed by reinforcement–decay dynamics (Agarwal et al., 19 Jun 2026). At query time, hierarchical intent gating decides whether dynamic knowledge is needed; if so, a salience-weighted semantic reranker scores nodes as $(1,52)$6, and a deterministic Context Card is injected into the system prompt with less than 10 ms overhead (Agarwal et al., 19 Jun 2026). In a 1,000-query world-lab simulation, PulseCX achieved IRR 89.2%, s-CSAT 4.2/5, and ER 13.5%, compared with Static RAG at 64.1%, 2.4/5, and 35.2%, and a Naive Online Agent at 78.5%, 3.1/5, and 18.9% (Agarwal et al., 19 Jun 2026). Here the “pulse” is the live external state of the world rather than a biosignal.

5. Adaptive hardware, structured uncertainty, and pulse-driven optimization

In some papers, AI on the pulse means closing the loop between sensing and hardware control. “Pulse excitation mode selection via AI Pipeline to Fully Automate the WUCT System” (Kumar et al., 2024) does this for water-coupled ultrasound computed tomography. The paper’s core hypothesis is that pulse width of the electrical excitation affects the emitted ultrasonic waveform, frequency content, attenuation pattern, and hence reconstruction quality (Kumar et al., 2024). The system uses a Random Forest Classifier trained on five line-scan features—total number of maxima, total number of minima, average signal value, maximum value, and minimum value—to predict the optimal pulse width from a discrete search over $(1,52)$7 ns (Kumar et al., 2024).

The learned optimizer is embedded in a larger automation stack that includes Intelligent Object Placement, data-quality assessment via Kanpur Theorem 1, and U-Net-based segmentation for reconstruction-quality scoring (Kumar et al., 2024). The segmentation model is trained on about 500 images and reaches validation segmentation accuracy 95.72% and IoU 0.8842 (Kumar et al., 2024). For the homogeneous Multi-Sample phantom, the RFC predicts pulse width 300 ns as best for 56% of line scans, followed by 250 ns for 20% and 350 ns for 13%; for the heterogeneous Rubber3M phantom, 250 ns, 300 ns, and 350 ns are each often optimal, suggesting that multiple pulse widths capture complementary structure (Kumar et al., 2024). This suggests a broader meaning of AI on the pulse: AI that chooses the pulse itself as a controllable acquisition parameter.

In rare-event physics, the CUORE pulse-shape dataset provides another angle (Collaboration, 25 Jun 2026). The data release contains 10,000 thermal pulses from a cryogenic calorimeter, equally split between 5,000 clean single pulses and 5,000 pile-up events, each as a 10,000-sample waveform at 1 kHz over a 10 s window (Collaboration, 25 Jun 2026). The normalized waveform $(1,52)$8 with $(1,52)$9 is recommended for supervised ML (Collaboration, 25 Jun 2026). This work does not report benchmark model performance, but it frames “AI on the pulse” as direct access to raw, labeled pulse time series for testing supervised and unsupervised methods in a high-stakes experimental setting.

Similarly, in materials science, PULSE becomes an AI-native variational estimator of partition functions rather than a signal classifier (Bernard et al., 27 May 2026). The target distribution $_2$0 is learned with an inverse VAE using the Gumbel–Softmax trick, and the model estimates observable-weighted partition functions $_2$1 from which thermodynamic averages are recovered (Bernard et al., 27 May 2026). On 2D Ising benchmarks, the method estimates canonical and observable-weighted partition functions with less than 1% relative error on small lattices while using orders of magnitude fewer samples than exhaustive enumeration (Bernard et al., 27 May 2026). This suggests that “pulse” can also be metaphorical, referring to the central thermodynamic “heartbeat” of the system via its partition function.

6. Collaboration, interpretability, and recurrent tensions

Across these works, several recurring design tensions define the technical meaning of AI on the pulse. One is the shift from static prediction to adaptive interaction. In PulseSatellite, analysts’ corrections trigger immediate fine-tuning (Logar et al., 2020); in PULSE for clinical endocrinology, iterative diagnosis generation is coupled to PubMed retrieval and re-reasoning (Huang et al., 11 Mar 2026); in PULSE for cancer survivorship, agentic LLMs choose which sensor streams and time windows to inspect, using eight MCP tools and per-user memory (Wang et al., 17 May 2026). These systems do not simply classify inputs; they continuously revise their internal understanding as more evidence arrives.

A second tension is between personalization and generalization. The home-monitoring anomaly detector uses per-patient fine-tuning, patient-specific thresholds, and reconstruction-based deviation scoring (Gabrielli et al., 5 Aug 2025). Pulse Focus uses normalization anchored to a normative sample but is intended to support personalized physiological or interaction models downstream (Debele et al., 2 Jun 2026). PULSE for survivorship explicitly compares current behavior against personal baselines while also retrieving peer cases for calibration (Wang et al., 17 May 2026). This suggests that “on the pulse” often entails maintaining a local baseline rather than relying on a population threshold.

A third is between interpretability and model expressiveness. The Acoustic Index is explicitly framed as a physics-informed, interpretable biomarker because it exposes Koopman modes, eigenvalues, and attention weights (Begiashvili et al., 17 Jul 2025). The PPG cross-attention model argues that attention maps offer a step toward explainability by showing how motion signals modulate PPG-based HR estimation (Kasnesis et al., 2022). The AI-enhanced stethoscope paper does not include explainability modules and lists interpretability as a limitation (Ghouse et al., 18 May 2025). PULSE for endocrinology addresses this partly through literature-grounded reasoning and structured differential lists, but also raises automation-bias concerns when clinicians defer too strongly to the AI (Huang et al., 11 Mar 2026).

A fourth is the management of temporal uncertainty and staleness. PulseCX solves this through explicit reinforcement–decay and volatility classes (Agarwal et al., 19 Jun 2026). PulseSatellite solves it operationally by adapting models during the same response window (Logar et al., 2020). VPuRD solves it by learning a virtual diagnostic that predicts the unmeasured lasing-off state from contemporaneous machine signals (Korten et al., 2024). In each case, the key problem is not merely noisy measurement but the instability of the underlying world between observation and action.

Finally, there is a common distinction between feasibility and deployment maturity. The AI-enhanced stethoscope reports a technical system with 94% classification accuracy but no large-scale clinical trial (Ghouse et al., 18 May 2025). The home-monitoring anomaly detector reports real deployment and clinician confirmation but only on 6 patients (Gabrielli et al., 5 Aug 2025). PULSE for clinical diagnosis reaches expert-competitive accuracy on 82 endocrinology cases yet explicitly documents automation-bias risks and the limits of vignette-based evaluation (Huang et al., 11 Mar 2026). PulseCX reports large gains in a simulated world-lab setting rather than production deployment (Agarwal et al., 19 Jun 2026). This suggests that AI on the pulse is as much a systems-design orientation as a completed application domain.

In sum, AI on the pulse is best understood not as a single method but as a technical paradigm for aligning AI with signals, behaviors, and world states that evolve on operational time scales. It includes low-cost auscultation and wearable anomaly detection (Ghouse et al., 18 May 2025, Gabrielli et al., 5 Aug 2025), pulse-signal fusion and CPR-state inference (Kasnesis et al., 2022, Sashidhar et al., 2020), virtual diagnostics and pulse optimization in physical systems (Korten et al., 2024, Krakowski et al., 2023), adaptive imaging and deployment-aware agent evaluation (Logar et al., 2020, Gao et al., 27 Apr 2026), and dynamic memory architectures for real-time external grounding (Agarwal et al., 19 Jun 2026). A plausible implication is that the unifying scientific contribution of this literature is architectural rather than purely algorithmic: AI systems become “on the pulse” when they are designed to sense continuously, represent temporality explicitly, personalize against changing baselines, and adapt their reasoning or control loops without waiting for a static benchmark or manually curated update cycle.