ShifaMind: Interpretable ICD-10 Coding

Change the form of the bottleneck, not its width

Concept Bottleneck Models make predictions auditable by routing them through human-readable concepts, but squeezing a rich clinical representation through a narrow scalar concept layer chokes capacity. Prior variants fix this by widening the bottleneck, which costs the inspectable interface. ShifaMind keeps the scalar concept interface and instead changes the bottleneck's form.

Concept grounding

Learnable concept queries attend to the note, producing a full-dimensional concept-grounded representation that is concept-relevant, but not compressed to 160 scalars.

Multiplicative gate

A learned gate g = σ(·) modulates that representation element-wise. The encoder summary can only steer which concept dimensions matter; it never bypasses them.

No-bypass guarantee

There is no direct path from the encoder to the diagnosis head. If the concept representation is zero, the output collapses to a constant, so prediction is structurally concept-mediated.

Results on MIMIC-IV top-50

Seven models on the same MIMIC-IV v3.1 top-50 split (79,742 / 17,088 / 17,088, seed 42), one shared global threshold τ = 0.5, no per-model tuning. Best per column in brass; ShifaMind row highlighted.

Model	Mac-F1	Mic-F1	Mac-AUC	Mic-AUC	P@5	P@8	P@15	R@5	R@8	R@15
CAML	0.674	0.716	0.939	0.954	0.633	0.508	0.325	0.698	0.833	0.942
LAAT	0.711	0.746	0.947	0.961	0.654	0.521	0.329	0.718	0.850	0.950
PLM-ICD	0.650	0.699	0.929	0.947	0.623	0.495	0.317	0.686	0.814	0.925
KEPT	0.687	0.727	0.941	0.957	0.643	0.508	0.323	0.706	0.833	0.939
GKI-ICD	0.649	0.700	0.930	0.949	0.626	0.496	0.318	0.689	0.816	0.928
Vanilla CBM	0.164	0.372	0.699	0.779	0.374	0.317	0.235	0.396	0.528	0.708
ShifaMind	0.712	0.735	0.942	0.959	0.647	0.517	0.327	0.711	0.845	0.947
Gemini 2.5 Prozero-shot*	0.435	0.547	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a
GPT-5.4zero-shot*	0.417	0.561	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a
Claude Sonnet 4.6zero-shot*	0.343	0.475	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a

* Zero-shot prompt evaluation on the same MIMIC-IV top-50 test set, with no fine-tuning. Reported for context only: these models are not optimized on the task, so this is not an apples-to-apples comparison with the trained models above. AUC and ranking metrics (P@K, R@K) do not apply to a zero-shot prompt that returns a code set rather than calibrated probabilities or a ranked list. 95% CIs on Macro-F1: Gemini 2.5 Pro [0.370, 0.455], GPT-5.4 [0.364, 0.441], Claude Sonnet 4.6 [0.300, 0.367].

ShifaMind (ours) Best in column Zero-shot LLM (context)

Statistical significance: paired bootstrap on Macro-F1

Comparison	F1 (ours)	F1 (other)	Δ	95% CI	p
vs. LAAT	0.712	0.711	+0.001	[−0.001, +0.003]	0.452
vs. KEPT	0.712	0.687	+0.025	[+0.022, +0.028]	<10⁻⁴
vs. CAML	0.712	0.674	+0.038	[+0.036, +0.041]	<10⁻⁴
vs. PLM-ICD	0.712	0.650	+0.062	[+0.059, +0.065]	<10⁻⁴
vs. GKI-ICD	0.712	0.649	+0.064	[+0.060, +0.067]	<10⁻⁴
vs. Vanilla CBM	0.712	0.164	+0.548	[+0.545, +0.552]	<10⁻⁴

Statistically indistinguishable from LAAT (the strongest baseline) on point accuracy, and significantly ahead of every other model, while being the only one that produces concept-mediated explanations.

Performance across code frequency: HEAD / MID / TAIL

Model	HEAD (16)	MID (16)	TAIL (18)	Overall
CAML	0.711	0.671	0.644	0.674
LAAT	0.740	0.715	0.682	0.711
PLM-ICD	0.699	0.658	0.600	0.650
KEPT	0.722	0.699	0.645	0.687
GKI-ICD	0.696	0.658	0.598	0.649
Vanilla CBM	0.341	0.124	0.042	0.164
ShifaMind	0.742	0.722	0.677	0.712

Best on HEAD and MID codes, competitive on the rare TAIL. the gain is not driven only by frequent labels.

Ablation: what each component contributes

Configuration	Macro-F1	Macro-AUC	Δ F1
ShifaMind (full)	0.712	0.942	ref.
w/o alignment loss	0.715	0.945	+0.003
w/o cross-attention (p_c ← p_t)	0.689	0.936	−0.023
Vanilla CBM (additive bottleneck)	0.164	0.699	−0.548

Removing concept grounding (cross-attention) costs real performance; the multiplicative form vs. a matched additive bottleneck is the difference between 0.712 and 0.164.

Precision / recall at τ = 0.5

Model	Macro-P	Macro-R	Micro-P	Micro-R
LAAT	0.757	0.683	0.778	0.716
KEPT	0.734	0.658	0.761	0.696
CAML	0.725	0.653	0.748	0.686
GKI-ICD	0.719	0.610	0.752	0.654
PLM-ICD	0.710	0.615	0.746	0.657
Vanilla CBM	0.348	0.162	0.415	0.338
ShifaMind	0.674	0.761	0.680	0.800
Gemini 2.5 Pro*	0.476	0.487	n/a	n/a
GPT-5.4*	0.411	0.517	n/a	n/a
Claude Sonnet 4.6*	0.309	0.450	n/a	n/a

At a shared threshold, ShifaMind sits at a recall-favoring operating point; LAAT at a precision-favoring one. * zero-shot macro precision / recall (micro not computed).

Interpretability isn't a claim, it's measured

Against a capacity-matched Vanilla CBM (same backbone, context, optimizer, loss, where only the bottleneck differs), ShifaMind wins on three behavioral metrics with non-overlapping bootstrap 95% confidence intervals.

CSTPR

Concept-Supported True Positive Rate: are correct diagnoses backed by relevant concepts?

ShifaMind0.704

Vanilla CBM0.147

4.8× higher

CIM

Concept Influence Magnitude: how strongly the concept representation drives the output.

ShifaMind1.314

Vanilla CBM0.645

2.0× at the diagnosis head

CCR

Concept-Conditioned Recall: when a relevant concept is present, is the diagnosis recovered?

ShifaMind0.836

Vanilla CBM0.361

83.6% vs 36.1%

A complementary behavioral test confirms it: masking the token spans of a diagnosis's top concepts drops that diagnosis's probability by 0.114 more than other diagnoses in the same note (95% CI [0.103, 0.127]). Predictions are sensitive to the concepts the model points to.

The architecture, end to end

A discharge summary becomes an inspectable, concept-mediated set of ICD-10 codes in four stages.

Long-context encoding

BioClinical ModernBERT-base encodes the full discharge summary (up to 8,192 tokens) into token representations and a p_t summary.

Concept grounding

160 learnable concept queries cross-attend to the note, producing a full-dimensional concept-grounded representation p_c, one query initialized per named clinical concept.

Multiplicative bottleneck

A gate conditioned on p_t and p_c modulates p_c element-wise: z = LN(g ⊙ p_c). No path bypasses the concept representation.

Diagnosis + inspection

The diagnosis head predicts multi-label ICD-10 codes from z. A separate concept head exposes scalar concept activations for clinician inspection, supervised via NegEx pseudo-labels.

A multiplicative concept bottleneck for interpretable ICD-10 coding.

Change the form of the bottleneck, not its width

Concept grounding

Multiplicative gate

No-bypass guarantee

Results on MIMIC-IV top-50

Interpretability isn't a claim, it's measured

The architecture, end to end

Long-context encoding

Concept grounding

Multiplicative bottleneck

Diagnosis + inspection

What we don't claim

Cite & explore