PRISM-Consult: Early Diagnostic Support

• PRISM-Consult is a clinician-aligned diagnostic system designed for early ED assessment of chest-pain-like cases using structured EHR data.
• It extends the compact PRISM transformer with a router and domain-specific LoRA adapters, ensuring parameter efficiency, low latency, and auditability.
• The framework employs a safety-first routing strategy with calibrated thresholds that prioritize life-threatening conditions and facilitate clinician oversight.

PRISM-Consult is a clinician-aligned panel-of-experts architecture for early diagnostic support in the Emergency Department (ED), especially for chest-pain-like presentations whose differentials span cardiac, pulmonary, gastro-oesophageal, musculoskeletal, and psychogenic causes. It extends the compact PRISM sequence model into a routed family of domain specialists: episodes are tokenized as structured clinical events; a light-weight router reads the first few tokens and dispatches to specialist models; and each specialist inherits PRISM’s small transformer backbone and token template, enabling parameter efficiency, low latency, and auditability (Levine et al., 1 Oct 2025).

1. Clinical objective and scope

PRISM-Consult is designed for ED use with the earliest, sparsest part of the encounter, including triage complaints, first labs, and early diagnostic orders. Its stated goal is to predict likely next events and diagnostic trajectory, and to surface clinically aligned differentials and diagnoses early, while remaining fast, auditable, interpretable, and conservative for life-threatening conditions (Levine et al., 1 Oct 2025).

The framework assumes structured EHR data from MIMIC-IV, restricted to demographics, admission and discharge events, diagnostic orders, labs, and diagnoses; procedures and medications are intentionally excluded. Each episode is an admission-to-discharge timeline. For routing, only the first $K$ events are used, with default $K=5$. Labels for specialist training are final ICD-9 diagnosis families, while router supervision uses early ICD-9 “initial” codes together with the final domain assignment (Levine et al., 1 Oct 2025).

The clinical decomposition follows five organ-system-aligned specialists: Cardiac–Vascular, Pulmonary, Gastro–Oesophageal, Musculoskeletal, and Psychogenic. The paper frames this decomposition around chest-pain-like presentations in which acute coronary syndromes, pulmonary embolic or pleural processes, gastro-oesophageal causes, chest-wall disorders, and panic or anxiety syndromes coexist in routine ED differentials. This suggests that PRISM-Consult is meant less as a generic triage engine than as a narrowly scoped, high-frequency diagnostic consult layer for a specific ED presentation family (Levine et al., 1 Oct 2025).

2. Architectural organization

The base PRISM model is a small decoder-only transformer trained as an autoregressive event sequence model. Its architecture uses $L = 6$ transformer blocks, model dimension $d_{\text{model}} = 256$, 4 attention heads, MLP expansion $4d$, learned absolute positional embeddings, and tied input-output token embeddings. The core language-modeling objective is

$$\mathcal{L}_{\text{LM}}(\theta) = -\sum_{t=1}^{T} \log p_\theta(x_t \mid x_{<t}),$$

with an auxiliary time-to-next-event regression head using Huber loss during training only (Levine et al., 1 Oct 2025).

PRISM-Consult wraps this backbone in a router-plus-specialists design. The router reads the first few events and outputs calibrated probabilities over the five domains. The specialists are five PRISM backbones adapted to their domains using LoRA adapters. Arbitration merges predictions when multiple specialists are consulted, using the fixed priority ordering Cardiac $>$ Pulmonary $>$ Gastro $>$ Musculoskeletal $>$ Psychogenic (Levine et al., 1 Oct 2025).

Parameter sharing is central. All specialists share the tokenization schema, embedding matrix, and the core 6-layer transformer parameters from the base PRISM model. Domain-specific adaptation is limited to LoRA adapters inserted into attention and feedforward layers, so only low-rank domain parameters are trained while shared PRISM weights stay frozen. This arrangement is explicitly presented as the source of parameter efficiency and low-latency consult-at-scale behavior (Levine et al., 1 Oct 2025).

3. Event representation and tokenization

PRISM-Consult represents an episode as a sequence

$K=5$0

where each event is

$K=5$1

Event sources include demographics, admission and discharge events, diagnostic orders, lab observations with discrete bins, and diagnoses (Levine et al., 1 Oct 2025).

Timestamps are normalized to time since admission, events are sorted chronologically, and ties are broken by event type priority DIAG $K=5$2 LAB $K=5$3 ORDER and then alphabetically. Time-gap markers such as $K=5$4 are inserted when inter-event intervals exceed thresholds. The token templates are explicit and semantically typed: diagnoses become $K=5$5, labs become $K=5$6, orders become $K=5$7, and special tokens include $K=5$8, $K=5$9, $L = 6$0, and $L = 6$1 (Levine et al., 1 Oct 2025).

Lab values are discretized into clinically meaningful bins LOW, NORMAL, HIGH, and CRITICAL. Missingness is handled implicitly: if no event exists, no token is produced. Sequences are truncated at 512 tokens per patient. The paper treats this event-template design as an interpretability feature, because tokens remain clinically readable and comparable across domains (Levine et al., 1 Oct 2025).

4. Router model and safety-first dispatch

The router consumes the first $L = 6$2 tokens of an episode. These tokens are concatenated into a short document, transformed into 1–2-gram TF–IDF features, and reduced by truncated SVD to 256 dimensions,

$L = 6$3

with optional time features concatenated to form $L = 6$4 (Levine et al., 1 Oct 2025).

The final router is a set of one-vs-rest logistic regression heads, one per domain, each calibrated with Platt scaling. For domain $L = 6$5,

$L = 6$6

The resulting calibrated probability vector is

$L = 6$7

Training uses binary cross-entropy over domains,

$L = 6$8

with prefix expansion so that prefixes of lengths $L = 6$9 inherit the episode label and may be sample-weighted by $d_{\text{model}} = 256$0 (Levine et al., 1 Oct 2025).

Dispatch follows a safety-first policy. Let $d_{\text{model}} = 256$1 and define life-threatening domains $d_{\text{model}} = 256$2. With thresholds $d_{\text{model}} = 256$3 and $d_{\text{model}} = 256$4, routing proceeds as follows: if a life-threatening domain exceeds the high threshold, route to that top-1 specialist; else if any domain exceeds the lower threshold, route to top-2 domains; else fail open and consult all five specialists. The reported grid search selected

$d_{\text{model}} = 256$5

A further fail-open rule applies when $d_{\text{model}} = 256$6 or when vital signs cross hard-coded danger thresholds, in which case all five specialists are called (Levine et al., 1 Oct 2025).

5. Specialist training and inference behavior

Specialists are trained as domain-filtered autoregressive models. For domain $d_{\text{model}} = 256$7,

$d_{\text{model}} = 256$8

where $d_{\text{model}} = 256$9 denotes frozen shared PRISM parameters and $4d$0 the domain-specific LoRA parameters (Levine et al., 1 Oct 2025).

The training data distinguish initial from conclusive labels. Initial ICD-9 codes are the diagnoses and codes recorded before the first definitive diagnosis and are used to define router inputs and early prefixes, explicitly to avoid leakage of final labels. Conclusive labels are final discharge diagnoses grouped into the five domain families and are used as specialist gold labels (Levine et al., 1 Oct 2025).

The study reports 20,436 ED episodes total for specialist training, with overlapping domain pools, and 13,801 unique episodes for the router. Router domain counts are Cardiac 1,128; Pulmonary 3,150; Gastro–Oesophageal 4,500; Musculoskeletal 523; and Psychogenic 4,500. Splits are stratified at the patient level with a 70/10/20 train/dev/test partition (Levine et al., 1 Oct 2025).

Inference is deliberately conservative. If one specialist is chosen, it runs alone; if two are chosen, their outputs are merged under the fixed priority ordering; if confidence is low or danger rules fire, the system consults all five. The paper presents this as a practical compromise between compute savings and clinical risk management (Levine et al., 1 Oct 2025).

6. Empirical results

The specialist models show smooth convergence across all domains. Reported validation losses decrease from 1.8397 to 0.7917 for Cardiac–Vascular, 1.4753 to 0.7041 for Pulmonary, 1.1303 to 0.7004 for Gastro–Oesophageal, 2.6927 to 1.2692 for Musculoskeletal, and 0.8910 to 0.6289 for Psychogenic. Corresponding development perplexities are approximately 2.21, 2.02, 2.01, 3.56, and 1.88, respectively. The paper attributes the higher Musculoskeletal perplexity to data scarcity and broader etiologic heterogeneity (Levine et al., 1 Oct 2025).

The router achieves strong routing quality on the held-out test set. Recall$4d$1, defined as whether the routed set includes at least one correct domain, is 1.000 on both development and test. Recall$4d$2, defined as whether the routed set includes all correct domains, is 0.945 on development and 0.942 on test. Life-threat recall, restricted to Cardiac or Pulmonary episodes, is 0.973 on development and 0.965 on test (Levine et al., 1 Oct 2025).

The compute benefit is substantial. The expected number of consulted experts per episode is 1.565, corresponding to approximately 69% fewer expert calls than a consult-all policy over five specialists. Estimated end-to-end latency is 88.4 ms on development and 89.5 ms on test (Levine et al., 1 Oct 2025).

The paper repeatedly emphasizes per-domain reporting rather than aggregate dominance by common classes. This suggests that PRISM-Consult is intended to be evaluated as a clinical routing system whose failure modes matter unevenly across specialties, rather than as a single-score classifier (Levine et al., 1 Oct 2025).

7. Safety, auditing, and deployment considerations

Safety is implemented through asymmetric thresholds for Cardiac and Pulmonary routing, fail-open behavior under low confidence, and a fixed arbitration priority that favors higher-risk organ systems. The paper argues that this routing policy is more appropriate for ED deployment than symmetric top-1 classification. It also states that future versions should tighten the policy to enforce hard life-threat recall constraints of at least 0.98 and, where needed, use isotonic calibration near decision boundaries (Levine et al., 1 Oct 2025).

Auditing is built into the representation and decision path. The system logs router logits, calibrated probabilities, thresholds, routing decisions including fail-open, specialist logits, and final arbitration. Because tokens remain clinically legible and the router operates on TF–IDF features over structured tokens, the decision process is positioned as inspectable by clinicians, quality-improvement teams, and regulators (Levine et al., 1 Oct 2025).

The main limitations are explicit. The current router cohort is effectively single-label, so multi-label arbitration remains under-exercised. Class imbalance is marked, especially for Musculoskeletal cases. Learned absolute positional encodings limit context length. Most importantly, the reported test life-threat recall of 0.965 falls slightly short of the stated target of 0.98, which is why the paper outlines external and temporal replication, asymmetric life-threat thresholds, and multi-label arbitration as prerequisites for prospective clinical deployment (Levine et al., 1 Oct 2025).

PRISM-Consult therefore occupies a specific design space: a compact, structured, clinician-aligned routing architecture for early ED diagnostic support, combining shared autoregressive event modeling with domain specialization, calibrated early-prefix dispatch, and safety-first fallback logic. Its empirical profile suggests that the main claimed advance is not simply diagnostic accuracy, but a practical route to safe, auditable, and low-latency consult at scale (Levine et al., 1 Oct 2025).