QuanTiMedAI: Quantum-Enhanced Time-Series Model guided by Agentic AI for Cardiac Arrest Mortality Prediction

QuanTiMedAI: Quantum-Enhanced Time-Series Model guided by Agentic AI for Cardiac Arrest Mortality Prediction (arXiv:2608.06294v1) proposes a novel quantum-agentic framework that combines an agentic large language model (LLM) for clinically informed feature discovery with a compact quantum recurrent network for temporality aware mortality prediction. To the authors' knowledge, agentic LLM reasoning for clinical feature discovery and quantum recurrent modeling for temporal prediction have not previously been brought together within a single clinical decision-support pipeline.

The paper addresses a critical gap in cardiac arrest mortality prediction. The authors note that "despite the growing availability of electronic health record data, existing mortality prediction studies in this population largely depend on static summaries derived from early admission. Such approaches ignore the temporal progression of physiological deterioration and recovery that unfolds throughout a patient's ICU stay. They observe that most MIMIC-IV–based cardiac arrest mortality prediction models rely primarily on conventional machine learning approaches, including LASSO, XGBoost, and ensemble-based methods, which treat each patient as a fixed snapshot with no way to capture how a patient's condition evolves over time. Critically, the authors state: to the best of our knowledge, no prior study has investigated sequential time-series modeling for cardiac arrest mortality prediction on MIMIC-IV."

The paper identifies two clear gaps: (1) no prior study identified in our review has applied any sequential time-series model to in-hospital mortality prediction in cardiac arrest patients on MIMIC-IV, and (2) even though QLSTM has consistently outperformed classical LSTM on time-series data in multiple domains, it has never been applied to sequential clinical EHR data for any patient outcome prediction task. Additionally, the authors criticize conventional feature selection methods as purely data-driven with limited capacity to incorporate existing clinical knowledge about cardiac arrest physiology, whereas agentic AI enables iterative refinement, allowing the model to evaluate performance, reflect on outcomes, and revise feature selections accordingly.

The study uses the MIMIC-IV database, a publicly available de-identified electronic health record repository from Beth Israel Deaconess Medical Center. Patients were identified by cardiac arrest diagnosis codes (ICD-9 code 427.5 and ICD-10 codes prefixed with I46), restricted to adults (≥18 years) with a linked ICU stay; only the first admission and first ICU stay were retained. The prediction task was binary in-hospital mortality classification using clinical data collected exclusively within the first τ = 24 hours of ICU admission. The cohort comprised 2,307 patients, of whom 1,296 (56.2%) died during hospitalization. Statistically significant baseline differences between mortality groups included age (SMD = 0.199, p < 0.001), prior stroke (SMD = 0.105, p = 0.016), hypertension, congestive heart failure, and myocardial infarction.

Feature selection was performed using an agentic framework driven by a large language model (LLM), accessed locally via Ollama, specifically Gemma 4(E4B) with thinking mode enabled. The agent operates in two phases: initial feature selection and iterative performance-guided refinement, both restricted to the training set. At each configuration (K, T), the agent receives "a mortality-stratified statistical summary of the training set comprising feature-level correlation with the outcome, mean ± SD stratified by mortality status, temporal trend direction across the first and second halves of the observation window, and percentage missingness. The agent iteratively refined selections over R = 5 rounds, required to swap at least α = 0.10 fraction of features per round, with feedback including fold-mean AUROC, AUPRC, cross-entropy loss, and per-feature permutation importance scores" from a V = 10-fold stratified cross-validation protocol using an LSTM proxy network.

The selected features and LLM-derived weights construct an engineered severity score channel appended to the input tensor at each temporal step: s(t) = σ(Σ wixi(t) + b), where σ is the sigmoid function. This provides the downstream model with a clinically informed summary of patient deterioration at each time step, embedding domain knowledge directly into the input representation.

The proposed QuanTiMedAI architecture replaces each classical gate with a hybrid quantum-classical block consisting of a classical linear projection followed by a variational quantum circuit (VQC). The cell contains five VQCs distributed across the gate structure (forget, input, update/candidate, output, and hidden-state refinement), omitting the output-stage VQC from the original six-VQC Chen et al. specification. A key innovation is that a residual skip connection reinjects the raw input xt directly into the hidden-refinement VQC5, preserving clinical signal that would otherwise be lost through the quantum bottleneck. Each VQC operates on Q = 4 qubits with three stages: Hadamard-based data encoding with RY and RZ rotations, L = 2 variational layers with CNOT cyclic entanglement and Rot(φ1, φ2, φ3) gates, and Pauli-Z expectation measurement. All circuits were implemented and executed using the default.qubit statevector simulator in PennyLane, reflecting ideal simulation conditions.

Across all 54 (K, T) configurations (2 ≤ K ≤ 20, T ∈ 2, 4, 6, 8, 12, 24 h), QuanTiMedAI attains the highest mean AUROC (0.815), outperforming both the LSTM (0.810) and the random selection based LSTM (0.775). The single best configuration (K=18, T=2 h) reaches AUROC of 0.852 and AUPRC of 0.882, versus best LSTM (AUROC = 0.846 at K=15, T=24) and best random LSTM (AUROC = 0.833). Compared to recent MIMIC-IV cardiac arrest studies, QuanTiMedAI's AUROC of 0.852 represents an improvement of approximately 2.9% over the 0.828 reported by Jia et al. — the current state-of-the-art baseline.

The model operates with only 605 total parameters (120 quantum + 485 classical head), versus 281,729 trainable parameters for the classical LSTM at the same setting — a ≈466× reduction in parameters. Against the most meaningful comparison, a parameter-matched classical LSTM of 655 parameters, which achieves an AUROC of 0.835, QuanTiMedAI achieves 0.852 — a gain of 0.017 AUROC (a relative improvement of approximately 2%) that can be attributed to the quantum gating mechanism rather than model size.

A structured ablation across six architectural variants at K=18, T=2 found:

• QuanTiMedAI (5-VQC with re-injection): AUROC 0.852 (highest)

• 6-VQC with skip connection: 0.848

• 5-VQC without skip (input re-injection): 0.839

• Mean-pool readout (vs last-step): 0.838

• Chen et al. [7] original 6-VQC baseline: 0.838

• Parameter-matched classical LSTM (655 params): 0.835

• Full classical LSTM (281,729 params): 0.828

The authors report that xt re-injection contributes a 0.013 AUROC improvement (approximately 1.5% relative), and that QuanTiMedAI also surpasses the large LSTM by a margin of 0.024 AUROC. Statistical testing confirmed QuanTiMedAI significantly outperforms the parameter-matched classical LSTM (Bonferroni-corrected p = 0.018). The authors note the inversion where a model with more parameters performs worse is a classic signature of overfitting on the limited training set.

The Gemma-derived severity weights add interpretability: the largest positive coefficients belonged to lactate, anion gap, base excess, bicarbonate, GCS, and SpO2, matching established clinical knowledge regarding mortality risk. The top severity weights were lactate (0.65), base excess (0.55), anion gap (0.35), GCS (0.30), albumin (0.25), bicarbonate (0.20), pH and SpO2 (0.20/0.10), and systolic blood pressure (−0.15), with severity bias 0.50.

Bloch sphere analysis of the VQC showed that after entanglement and variational layers, variational layers leverage this space to repel states toward the boundaries, fracturing the data into two distinct, spatially segregated clusters, demonstrating the network distributes feature processing efficiently and confirming QuanTiMedAI successfully exploits Hilbert space expressivity for robust linear separability.

The authors conclude that QuanTiMedAI achieved "an AUROC of 0.852 using only 605 parameters, compared to 281,729 for a full-capacity LSTM, exceeding a parameter-matched classical LSTM (0.835) and improving on a recent MIMIC-IV cardiac arrest baseline by approximately 2.9%. These results support QuanTiMedAI as a lightweight candidate for future resource-aware clinical applications, pending external validation." The work's principal contributions are: (1) the quantum-agentic time-series framework integrating agentic LLM-guided feature selection, an engineered severity-score channel, and a compact quantum recurrent network; (2) the agentic LLM-guided feature selection pipeline with iterative refinement; (3) the modified five-VQC QLSTM architecture with input re-injection skip connection achieving 2% AUROC improvement over a parameter-matched classical LSTM; and (4) a rigorous six-variant ablation study isolating each design choice's contribution. The authors acknowledge that validation under realistic hardware noise remains an important next step, and that reported results reflect ideal simulation conditions and do not account for gate errors, decoherence, or readout noise present on physical quantum hardware.

Based on this paper, here are specific improvements to AI systems and what the improved systems can do:

1. Integrate agentic LLM-driven feature discovery into clinical time-series pipelines.

Instead of relying purely on data-driven feature selection (e.g., LASSO, permutation importance), the AI system can use an LLM agent that receives mortality-stratified statistics, temporal trends, missingness, and iterative performance feedback (AUROC, AUPRC, loss, permutation importance) to propose and refine feature sets over multiple rounds. The improved system can:

• Automatically incorporate clinical knowledge (e.g., lactate, anion gap, GCS, base excess) into feature selection without manual curation.

• Adapt feature choices based on model feedback, swapping at least 10% of features per round, leading to more robust and clinically grounded predictors.

2. Embed an LLM-derived severity score as a temporal input channel.

The AI system can compute a per-time-step severity score s(t) = σ(Σ wixi(t) + b), where weights and bias are generated by the LLM agent. This channel is appended to every temporal step of the input. The improved system can:

• Provide the downstream model with a compact, clinically meaningful summary of deterioration at each time step.

• Improve mortality prediction by injecting domain knowledge directly into the representation, even with very few parameters.

3. Use a compact quantum recurrent cell with residual input re-injection.

The AI system can replace classical LSTM gates with hybrid quantum-classical blocks (four-qubit variational circuits) and add a residual skip connection that reinjects the raw input xt into the hidden-refinement circuit. This improved system can:

• Achieve comparable or better AUROC (0.852) with only 605 parameters, a 466× reduction versus a full LSTM.

• Preserve clinical signal that would otherwise be lost through the quantum bottleneck, yielding ~1.5% relative improvement over a 5-VQC variant without the skip connection.

• Outperform a parameter-matched classical LSTM (0.835 → 0.852), demonstrating that the quantum gating mechanism itself contributes predictive power.

4. Design parameter-matched ablation protocols to isolate architectural contributions.

Instead of comparing only against full-size baselines, the AI system can use parameter-matched classical counterparts and a structured set of ablations (e.g., 5-VQC vs 6-VQC, with/without skip, mean-pool vs last-step readout, Chen et al. baseline). The improved system can:

• Determine whether performance gains come from model size, architectural specificities, or quantum components.

• Identify overfitting signatures (larger models performing worse) and guide resource-aware model selection.

5. Exploit Hilbert space expressivity for linear separability in time-series classification.

The AI system can intentionally use variational quantum circuits with entanglement and variational layers to repel latent representations toward decision boundaries. The improved system can:

• Achieve robust linear separability of mortality classes, as shown by Bloch sphere analysis.

• Distribute feature processing efficiently across quantum state space, enabling high performance at extremely low parameter counts.

6. Apply this quantum-agentic framework to resource-constrained and privacy-sensitive clinical settings.

Because the model runs on ~605 parameters and uses a local LLM via Ollama, the improved AI system can:

• Operate on edge devices or within hospital firewalls without sending EHR data to external clouds.

• Retain high predictive accuracy (AUROC 0.852) while requiring minimal memory and compute, facilitating real-time ICU monitoring and deployment in low-resource environments.

7. Iteratively refine feature sets using a proxy model during development.

The AI system can use a fast LSTM proxy with V-fold stratified cross-validation to generate feedback for the agentic LLM, rather than training the full quantum model each round. The improved system can:

• Reduce the computational cost of feature search while still converging to high-quality, clinically meaningful selections.

• Re-use the same refinement loop for other outcome prediction tasks (e.g., sepsis, acute kidney injury) by swapping the statistical summary and outcome label.

8. Replace static-summary mortality prediction with sequential modeling for time-evolving conditions.

The improved AI system can ingest multi-hour ICU time-series (e.g., 2–24 h windows) and predict in-hospital mortality using temporal progression rather than a fixed snapshot. The system can:

• Capture deterioration and recovery dynamics that static models miss, resulting in an AUROC improvement from 0.828 (state-of-the-art static baseline) to 0.852.

• Provide clinicians with dynamic risk scores that update as new hourly data arrive.