Skip to main content

​
Overview

HANA’s Deliberation Framework integrates multiple specialized clinical “expert” models to produce high-quality conversation plans through collaborative reasoning. Rather than relying on a single model to make clinical decisions, the framework orchestrates inputs from distinct, configurable specialty models — each evaluating the conversation plan from its domain perspective — to synthesize a unified, well-supported clinical approach. The deliberation framework is the reasoning engine behind HANA’s pre-call planning. It determines what questions to ask, in what order, with what branching logic, and when to escalate — before the conversation begins. This planning step is what enables the real-time conversation engine to operate within tight latency constraints while maintaining clinical accuracy.

​
Core Architecture

​
System Design

The Deliberation Framework comprises:
  • Triage Model: Identifies the clinical task type and selects appropriate specialty models for the conversation context
  • Expert Model Block: Consists of independent expert agents that evaluate tasks from specialized perspectives, each generating probability distributions over potential actions
  • Deliberation Aggregation Model: Synthesizes expert inputs into a final decision, leveraging nuanced evaluation and probabilistic weighting

​
Expert-Based Decomposition

The Triage Model evaluates clinical queries (Q) and assigns domain-specific models based on identified requirements: 
LLM_triage(Q) → (Task Type, Specialties)
Specialties = {s₁, s₂, ..., sₙ}
Each specialist model (LLM_sᵢ) generates an independent response and a probability distribution: 
R_expert^(i) = LLM_sᵢ(Q, Task Type)
The final aggregated output is determined through: 
O_final = f(R_expert^(1), ..., R_expert^(n))
A_final = LLM_consensus(O_final)

​
Probability Weighting

To reflect clinical uncertainty, the Deliberation Framework aggregates expert probabilities using a Weighted Log Opinion Pool (WLOP): 
P_combined(X) = Σ wᵢ · log pᵢ(X)
P_normalized(X) = exp(P_combined(X)) / Σⱼ exp(P_combined(Xⱼ))
This approach moderates expert biases, rewarding consensus while accommodating individual expert uncertainties.

​
Cascade Boosting for Probabilities

To further refine probability assessments, Cascade Boosting amplifies the likelihood of answers based on frequency and rank across experts:
  • Establish frequency f_x,r for answer X at rank r
  • Apply cascade weighting (θ) progressively decreasing for lower ranks
Boosted scores calculated: 
BoostedScore(X) = P_normalized(X) + λ_boost · Σ(f_x,r × θ_r)
The final boosted probability: 
P_final(X) = exp(BoostedScore(X)) / Σⱼ exp(BoostedScore(Xⱼ))

​
Reaching Consensus

The final decision integrates each expert’s analysis through structured deliberation, examining not only aggregated probabilities but also expert reasoning and justification. The Deliberation Model explicitly factors nuanced clinical reasoning, allowing a sophisticated, clinically coherent final recommendation beyond numerical probability alone.

​
Application to HANA Voice Conversations

​
Pre-Call Planning

The deliberation framework determines the conversation plan:
  1. Triage model analyzes patient profile and identifies clinical task type
  2. Expert ensemble evaluates proposed question set, ordering, and branching logic
  3. Consensus aggregation produces final conversation plan with confidence-weighted questions

​
Real-Time Branching

During conversation, branching decisions use a simplified version of the deliberation framework optimized for voice latency constraints:
  1. Patient response parsed and entities extracted
  2. Pre-computed branching rules (generated during pre-call planning) handle the majority of decisions with sub-second latency
  3. For unanticipated responses that fall outside pre-computed plans, a lightweight evaluation determines the optimal branch within the latency budget
  4. If no action achieves sufficient confidence, the system falls back to a safe default or escalates to staff
The full multi-model deliberation framework runs during pre-call planning (where the 3-second latency budget allows it). Real-time branching primarily relies on the pre-computed conversation plan, with lightweight fallback evaluation for edge cases.

​
Escalation Decisions

Deliberation governs escalation from AI to human:
  • Clinical escalation: patient reports symptoms requiring immediate attention
  • Emotional escalation: patient distress exceeds safe handling capacity
  • Complexity escalation: conversation diverged beyond protocol coverage
  • Quality escalation: system confidence dropped below acceptable threshold

​
Evaluation Methodologies

​
Accuracy Assessment

  • Raw accuracy: Overall accuracy measured with detailed breakdown by clinical specialty and conversation type
  • Top-K performance: Evaluates how frequently correct clinical actions appear within top-ranked probability outputs
  • Reliability analysis: Calibration assessments ensure confidence intervals reliably correlate with observed accuracy

​
Future Enhancements

  • Enhanced specialization: Expand domain-specific expert models for nuanced evaluation across diverse clinical domains
  • Predictive quality assessment: Implement proactive predictions about conversation quality to allow preemptive corrections
  • Cross-modal evaluation: Assess consistency across different clinical data modalities

​
Modular Framework

The Deliberation Framework is modular and adaptable:
  • Expert Model Blocks: Independently configurable expert models specialized in distinct clinical domains — allows targeted updates without disrupting overall system
  • Flexible Triage System: Dynamically routes clinical queries to appropriate expert models based on case specifics
  • Dynamic Deliberation Aggregation: Aggregates expert outputs using weighted consensus, customizable for accuracy, efficiency, or latency optimization
  • Real-Time Integration: Modular structure facilitates real-time adjustments based on clinician feedback and evolving clinical guidelines