MIMIC-IV v3.1 · top-50 · ranked by Macro-F1

Interpretable ICD-10 coding leaderboard

Seven models on one shared split (79,742 / 17,088 / 17,088, seed 42), a single global threshold τ = 0.5, no per-model tuning. ShifaMind leads while being the only entry that produces concept-mediated explanations.

#	Model	Macro-F1
1	ShifaMind 👑tied w/ LAAT, p=0.452	0.712
2	LAAT	0.711
3	KEPT	0.687
4	CAML	0.674
5	PLM-ICD	0.650
6	GKI-ICD	0.649
7	Gemini 2.5 Pro *	0.435
8	GPT-5.4 *	0.417
9	Claude Sonnet 4.6 *	0.343
10	Vanilla CBM	0.164

* Zero-shot LLMs: prompt-only, no fine-tuning on the task, shown for context. Read: ShifaMind is statistically tied with LAAT on point accuracy and significantly ahead of every other trained model, while being the only one with a concept-mediated, inspectable interface.

Macro-F1 0.712 160 clinical concepts BioClinical ModernBERT · 8,192 ctx

Read the preprint → Paper page on 🤗

Significance test · paired bootstrap on Macro-F1

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

The ShifaMind vs LAAT gap is a statistical tie at this seed; the separation from every other model is significant.

The idea

Why it leads without losing the audit trail

Concept Bottleneck Models stay auditable by routing predictions through human-readable concepts, but squeezing a rich clinical representation through a narrow scalar layer chokes capacity. ShifaMind keeps the scalar concept interface and changes the bottleneck's form instead of its width.

Concept grounding

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

Multiplicative gate

A learned gate g = σ(·) modulates that representation element-wise. The encoder summary steers which concepts matter; it never bypasses them.

No-bypass guarantee

No direct path from encoder to diagnosis head. Zero out the concept representation and the output collapses, so prediction is structurally concept-mediated.

Measured

Interpretability isn't a claim

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% CIs.

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%

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]).

Cite

Cite & explore

ShifaMind: A Multiplicative Concept Bottleneck for Interpretable ICD-10 Coding · Mohammed Sameer Syed, Xuan Lu · College of Information Science, University of Arizona.

BibTeX

@article{syed2026shifamind,
  title   = {ShifaMind: A Multiplicative Concept Bottleneck for Interpretable ICD-10 Coding},
  author  = {Syed, Mohammed Sameer and Lu, Xuan},
  journal = {arXiv preprint arXiv:2605.08482},
  year    = {2026}
}