DUAL-Health: Dual-Stream Health Modeling

• DUAL-Health is a dual-structured modeling approach that separately computes unsupervised anomaly detection and physics-based load estimation to capture complex health signals.
• It leverages explicit decomposition to isolate heterogeneous data and roles, enhancing monitoring, decision support, and system robustness.
• Applications span from vehicle health monitoring to digital twin systems and human-facing platforms, ensuring interpretable outcomes and improved safety.

DUAL-Health denotes a dual-structured approach to health-related modeling in which two complementary components are kept explicitly separate and then coordinated. In its most literal formulation, it is a “two-dimensional health state” in which an unsupervised anomaly score $A_{\text{ML}(t)}$ and a physics-based load score $W_{\text{Phys}(t)}$ are computed in parallel as $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$ for embedded vehicle monitoring (Spotorno et al., 11 Feb 2026). Across adjacent literatures, closely related designs recur as patient narrative versus structured clinical record, Observer versus Presenter, coach versus client simulator, and parent versus child use contexts (Pugh et al., 29 Apr 2026, Kovacevic et al., 16 Jun 2026, Long et al., 7 May 2026, Lee et al., 2023). This suggests that DUAL-Health is best understood as an architectural pattern for separating heterogeneous health signals, roles, or epistemic sources before fusing them for monitoring, decision support, or interaction.

1. Conceptual basis and architectural principle

The defining property of DUAL-Health systems is explicit decomposition. Rather than forcing one model, one state variable, or one user role to absorb all relevant variation, these systems preserve two coordinated channels whose failure modes are different. In the vehicle setting, the split is between “statistical anomaly detection (dynamic hazards)” and “physics-based cumulative load estimation (mechanical effort)” (Spotorno et al., 11 Feb 2026). In longitudinal coaching agents, it is between the “Narrative Stream” of patient self-report and the read-only “Clinical Stream” of FHIR-grounded record data (Pugh et al., 29 Apr 2026). In embodied reflection on wearable data, it is between an Observer agent that computes descriptive statistics and trends and a Presenter agent that turns those findings into “spoken statistics” (Kovacevic et al., 16 Jun 2026).

The same logic appears in dual-role interaction designs. A “dual dialogue system” for mental health places the AI in a back-channel visible only to the clinician, preserving therapist–client dialogue on one side and therapist–assistant interaction on the other (Kampman et al., 2024). In motivational-interviewing coaching, DACT formalizes a coach agent and a client simulator as two players in a stochastic game, with the coach maximizing multi-dimensional motivational-interviewing quality and the client implicitly maximizing challenge through preference reversal (Long et al., 7 May 2026). In pediatric mHealth, duality is sociotechnical rather than algorithmic: 30/43 apps are “explicitly or implicitly designed for dual-user scenarios,” 13/43 are effectively parent-only, and 0/43 are exclusively child-only (Lee et al., 2023).

Methodologically, this dual principle also appears in architectures that are not themselves named DUAL-Health. DuETT factorizes EHR modeling across “time and event type dimensions,” and semi-supervised segmentation via dual networks uses separate student and teacher subnetworks plus uncertainty-aware pseudo supervision (Labach et al., 2023, Lu et al., 23 May 2025). A plausible implication is that DUAL-Health belongs to a wider class of health AI systems that gain robustness by decomposing structure along clinically meaningful axes rather than flattening everything into a single latent stream.

2. Canonical formulation: dual-stream operational health monitoring

The most explicit DUAL-Health formulation appears in “A Dual-Stream Physics-Augmented Unsupervised Architecture for Runtime Embedded Vehicle Health Monitoring” (Spotorno et al., 11 Feb 2026). The system processes low-frequency $10\ \text{Hz}$ inertial and GNSS-derived signals $a_x(t), a_y(t), a_z(t), v(t)$, segmented into non-overlapping windows of 30 samples (3 s), and produces the health vector

$$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$

Here $A_{\text{ML}(t)}$ is the anomaly score from a symmetric LSTM autoencoder trained on smooth-road data, with encoder sizes $128 \rightarrow 64 \rightarrow 32$, Adam learning rate $\eta = 10^{-3}$, and reconstruction-based scoring

$$A_{\text{ML}(t)} = \sqrt{\frac{1}{N}\sum_{i=1}^{N}(x_i-\hat{x}_i)^2}.$$

This stream detects transient shocks, rough terrain, and highly unpredictable signals, but it gives low scores not only to benign cruising but also to “smooth, high-load operations” such as constant-speed hill climbs (Spotorno et al., 11 Feb 2026).

The second stream uses macroscopic physics proxies derived from the same windows. After pitch estimation, gravity compensation,

$W_{\text{Phys}(t)}$0

and jerk estimation by symmetric central difference with a 5-point moving average, it computes four load or stress proxies: suspension stress $W_{\text{Phys}(t)}$1, lateral stress $W_{\text{Phys}(t)}$2, drivetrain stress $W_{\text{Phys}(t)}$3, and braking stress $W_{\text{Phys}(t)}$4. The vibration-related proxies integrate squared jerk, while drivetrain stress is defined as positive tractive power,

$W_{\text{Phys}(t)}$5

and braking stress as dissipative energy accumulated during deceleration. These are aggregated into a normalized scalar $W_{\text{Phys}(t)}$6, yielding a dimensionless load-intensity score for each window (Spotorno et al., 11 Feb 2026).

Scenario	Stream A (ML)	Stream B (Physics)
Highway cruise	Low	Low
Uphill towing	Low	High
Pothole/shock	High	Low
Rough terrain	High	High

The table expresses the central blind spot that motivated DUAL-Health: “high-load steady states” can be mechanically harsh while appearing statistically normal. The paper therefore uses a max-pooling fusion

$W_{\text{Phys}(t)}$7

so that either instability or sustained work can trigger attention (Spotorno et al., 11 Feb 2026).

Validation used CARLA with a heavy-duty Firetruck model over 218 runs, vehicle masses 8300–13,500 kg, and $W_{\text{Phys}(t)}$8 windows. The physics proxies showed discriminative power and behaved as expected mechanical stress indicators: suspension stress distinguished potholes from normal driving with $W_{\text{Phys}(t)}$9 versus $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$0, drivetrain stress separated ramp climbs from level roads with $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$1 versus $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$2, and the mass–drivetrain-stress Spearman correlation was $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$3 (Spotorno et al., 11 Feb 2026). By contrast, the unsupervised anomaly stream correlated only weakly with drivetrain and braking stress, with $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$4 and $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$5, confirming structural orthogonality rather than simple undertraining (Spotorno et al., 11 Feb 2026).

A further feature of the canonical DUAL-Health design is deployability. On a StarFive VisionFive 2 (RISC-V JH7110, quad-core U74 @ 1.5 GHz), measured per 3 s window and single-threaded, Stream A required $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$6 and $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$7, Stream B required $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$8 and $$H_t = \big[A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big]$$9, and the total dual-stream cost was $10\ \text{Hz}$0 and $10\ \text{Hz}$1, with the physics stream adding approximately $10\ \text{Hz}$2 overhead (Spotorno et al., 11 Feb 2026).

3. Dual roles in human-facing health systems

In human-facing systems, DUAL-Health often appears as role separation rather than sensor fusion. A qualitative review of children’s health apps found that dual use is effectively the norm: 30/43 apps are explicitly or implicitly dual-user, 13/43 are parent-only in practice, and none are purely child-only (Lee et al., 2023). The study also identified five goals—“Instructional & educational,” “Tracking,” “Introspection,” “Communication,” and “Telehealth services”—and argued that design should make role expectations explicit, support graduated autonomy, and move beyond vague “for kids” framing. This establishes DUAL-Health, in the pediatric app setting, as explicit co-configuration of child-facing and caregiver-facing functions rather than a single undifferentiated interface (Lee et al., 2023).

Mental health support systems make the role split algorithmically explicit. A “multi-agent dual dialogue system” supports mental health professionals by placing the AI in a parallel channel that proposes responses, analyzes conversations, summarizes dialogue, and recommends localized psychoeducation and internet-based cognitive behavioral therapy exercises from a curated corpus (Kampman et al., 2024). The client sees only the therapist–client chat, while the therapist has a second pane for “Propose response,” “Recommend resources,” “Analyze conversation,” “Summarize conversation,” “Empathetic rewrite,” and “Open-ended chat.” On 100 query–response pairs from a Singapore mental-health forum, responses generated by GPT‑4o, Llama 3 8B, and Llama 3 70B were rated on an adapted Therapist Empathy Scale; the average human response score was $10\ \text{Hz}$3, GPT‑4o scored $10\ \text{Hz}$4, Llama 3 70B scored $10\ \text{Hz}$5, and Llama 3 8B scored $10\ \text{Hz}$6, with no significant overall average differences across sources (Kampman et al., 2024).

A simpler but closely related dual-agent design appears in wearable-data reflection. “Talking to Your Data” separates an Observer agent, which computes descriptive statistics, correlations, and temporal trends using weighted linear regression with decay, from a Presenter agent, which is strictly grounded to an Insight JSON and communicates findings through short, non-clinical “spoken statistics” (Kovacevic et al., 16 Jun 2026). In a within-subject simulated-self study with $10\ \text{Hz}$7, the embodied agent condition produced a median perceived actionability of 7.0 versus 6.0 for dashboards and increased the mean Specificity Score of generated actions from approximately 1.25 to approximately 2.0, while reported concentration dropped from a median of 5.0 to 2.0 (Kovacevic et al., 16 Jun 2026).

DACT generalizes this role duality into co-training. It models coach and client as separate policies in a stochastic game, uses a multi-dimensional LLM judge over “Cultivating Change Talk,” “Softening Sustain Talk,” and “Empathy,” and trains both agents with DPO on Pareto-dominant response pairs (Long et al., 7 May 2026). On held-out personas, DACT achieved a 4-condition average mean3 score of $10\ \text{Hz}$8 and anti-pattern rate of $10\ \text{Hz}$9, compared with $a_x(t), a_y(t), a_z(t), v(t)$0 and $a_x(t), a_y(t), a_z(t), v(t)$1 for GPT‑Coach, and $a_x(t), a_y(t), a_z(t), v(t)$2 and $a_x(t), a_y(t), a_z(t), v(t)$3 for SFT (Long et al., 7 May 2026). The pattern is consistent: DUAL-Health in interaction design separates roles so that support, challenge, and oversight are not collapsed into one agentic surface.

4. Dual sources of truth, multidimensional state, and digital twinning

A second major interpretation of DUAL-Health is epistemic separation. In longitudinal coaching agents, a Dual-Stream Memory Architecture maintains a patient-reported Narrative Stream $a_x(t), a_y(t), a_z(t), v(t)$4 and a read-only Clinical Stream $a_x(t), a_y(t), a_z(t), v(t)$5 derived from FHIR, with reconciliation applied to every new memory rather than overwriting the record with the latest utterance (Pugh et al., 29 Apr 2026). The reconciliation function

$a_x(t), a_y(t), a_z(t), v(t)$6

returns discrepancy detection, implicated FHIR resources, severity, safety-criticality, and justification. On 26 patients and 675 longitudinal wellness-coaching sessions, isolated testing detected 84.4% of designed clinical discrepancies with 86.7% safety-critical recall, while end-to-end testing with extracted memories produced a 13.6 percentage-point error cascade attributable to extraction losses rather than downstream classification (Pugh et al., 29 Apr 2026). The architecture formalizes a central DUAL-Health proposition: conflicting sources of truth should be reconciled, not silently merged.

Remote patient monitoring makes the same principle quantitative. A multidimensional RPM model represents patient health as

$a_x(t), a_y(t), a_z(t), v(t)$7

and couples it with monitoring state $a_x(t), a_y(t), a_z(t), v(t)$8, so that decisions are made over service states $a_x(t), a_y(t), a_z(t), v(t)$9 (Chandak et al., 4 Mar 2025). The optimal policy is characterized by “switching curves” or “switching hyper-surfaces”: intensive monitoring becomes optimal when the health vector crosses a multidimensional boundary. For critical sets of the form $$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$0, the paper shows that, for sufficiently small $$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$1, the optimal control has a threshold form with switching curve $$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$2 for some $$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$3 (Chandak et al., 4 Mar 2025). This is not termed DUAL-Health in the paper, but it is structurally aligned with the idea that health management should respond to multiple dimensions simultaneously rather than to a scalar proxy.

Digital-twin work extends the dual principle from state variables to ontological layers. MyDigiTwin combines a patient’s “real” side, aggregated through Personal Health Environments and ZIB-FHIR, with a “virtual” side implemented as a digital twin for privacy-preserving cardiovascular risk prediction and scenario exploration (Cadavid et al., 21 Jan 2025). Using federated learning with FedAvg, a deep neural Cox model trained on harmonized Lifelines data improved global C-statistic from 0.764 to 0.788, while preserving the principle that data remain local and only model parameters travel (Cadavid et al., 21 Jan 2025). OmniBioTwin generalizes this further into a “System-of-Twinned-Systems” with seven layers—Data, Twin, Coupling, Synchronization, Decision, Interaction, and Audit—and formal twin definition

$$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$4

with inter-twin coupling

$$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$5

and system-level decision

$$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$6

(Wang et al., 9 Jun 2026). A plausible implication is that DUAL-Health scales naturally from two streams to layered systems of coupled twins, provided the separations remain explicit and auditable.

5. Interpretability, evaluation, and safety calibration

DUAL-Health systems place unusual weight on interpretable intermediate artifacts. In suicidality analysis, a dual-prompting pipeline first performs knowledge-aware evidence extraction with MentaLLaMA-chat-13B, an expert identity prompt, and a suicide dictionary, then performs evidence summarization with SOLAR-10.7B and an LLM-based consistency evaluator (Jeon et al., 2024). On the CLPsych 2024 shared task, few-shot extraction improved recall from 0.910 to 0.922 and F1 from 0.911 to 0.917; for summarization, a SOLAR-only configuration achieved consistency 0.973, slightly above the 0.970 obtained when combining SOLAR and MentaLLaMA (Jeon et al., 2024). The dual structure here is interpretive: spans first, summary second.

In multi-modal mental-health understanding, ECMC separates emotion and cognition into dual-stream BridgeNets based on Q-former, producing embeddings $$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$7 and $$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$8 that condition a LLaMA decoder for emotion–cognition captioning (Zhou et al., 2 Mar 2026). The resulting emotion–cognition profiles improved downstream diagnosis, with average gains of $$H_t = \big[\,A_{\text{ML}(t)},\, W_{\text{Phys}(t)}\big].$$9 ACC and $A_{\text{ML}(t)}$0 F1 for depression, and $A_{\text{ML}(t)}$1 ACC and $A_{\text{ML}(t)}$2 F1 for anxiety, relative to raw baselines (Zhou et al., 2 Mar 2026). This again reflects a DUAL-Health pattern: clinically meaningful factors are decomposed first and fused later.

Safety evaluation introduces another duality: refusal versus helpful completion. Health-ORSC-Bench measures Over-Refusal Rate on 31,920 benign boundary prompts and Safe Completion Rate

$A_{\text{ML}(t)}$3

where $A_{\text{ML}(t)}$4 comprises “Partial Answer” and “Full Answer” labels for safe responses (Zhang et al., 25 Jan 2026). Across 30 models, safety-optimized systems frequently refused up to 80% of “Hard” benign prompts, while models such as Qwen‑Max and Qwen‑Plus showed near-zero over-refusal but weaker toxic rejection (Zhang et al., 25 Jan 2026). In health contexts, DUAL-Health thus becomes a calibration problem: maximizing safety without collapsing benign, ambiguous, or research-oriented requests into blanket refusal.

ALPHA provides a physiologic example of dual-role evaluation. It treats LLMs as both analyzers of physiological data and user-facing health assistants, reporting mean absolute error below 1 beat per minute for heart rate and below 1% for oxygen saturation, with overall health-assessment accuracy surpassing 85% on anomalous plateau data, while a vision-enabled GPT on PPG achieved less than 1 bpm error in cycle count and 7.28 MAE for heart-rate estimation (Tang et al., 2023). The paper explicitly describes this as a “dual role,” reinforcing the view that DUAL-Health concerns not only multiple streams of data but also multiple functional roles of AI within one health system.

6. Limitations and future trajectory

The literature is consistent in identifying limits of current DUAL-Health instantiations. The vehicle architecture is validated on simulated CARLA data rather than real vehicles; it uses 10 Hz sampling that misses high-frequency micro-vibrations, relies on rough mass estimates and simplified drag modeling, and does not yet provide explicit component-level damage models (Spotorno et al., 11 Feb 2026). In pediatric mHealth, a search explicitly including chronic-disease terms still found no widely used chronic illness management apps for children that met the inclusion criteria, indicating either scarcity or low adoption, while most child-facing apps still ignore fine-grained age segmentation (Lee et al., 2023).

Longitudinal conversational systems face different limits. The dual dialogue system for mental health has not yet been tested in real clinical settings, raises privacy concerns around commercial APIs, and must contend with hallucinations, cultural bias, and automation bias (Kampman et al., 2024). The dual-stream memory architecture shows that reconciliation is feasible, but also that clinical safety depends critically on extraction fidelity: the main failure mode is the loss of clinical detail during memory extraction from unstructured conversation (Pugh et al., 29 Apr 2026). Multidimensional RPM models, although structurally interpretable, face the curse of dimensionality, assume Markovian dynamics and full observability, and omit explicit capacity constraints (Chandak et al., 4 Mar 2025).

At larger scale, system-of-twins architectures remain largely programmatic. OmniBioTwin identifies unresolved challenges in multimodal alignment, biologically grounded coupling operators, uncertainty propagation, empirical validation, and governance (Wang et al., 9 Jun 2026). This suggests that the next phase of DUAL-Health research will depend less on introducing new forms of duality than on stabilizing existing ones: real-world validation, uncertainty-aware coupling, role-explicit interfaces, privacy-preserving deployment, and audit-ready decision support.