acceptodds
Withdrawn from ICLR 2027

TheraRoute-L: Compositional Treatment-Unit Learning for Multimodal Response Prediction across Clinical Settings

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Coarse treatment encodings leave drug composition and order implicit in multimodal prediction of pathological complete response (pCR) to neoadjuvant therapy. Learning patient-dependent interactions is further complicated by variable imaging histories and missing inputs. We introduce TheraRoute-L, jointly learning record-supported treatment-unit and relation representations under pCR supervision while retaining whole-regimen semantics. Clinical fields and observed MRI histories condition unit reads and bilinear scores of concurrent or sequential relations before fusion. A causal history encoder updates patient state, and seven private modality-subset experts combine available predictions. We derive interaction and read-sensitivity bounds and pooled and direct-read gradient paths, showing how direct reads bypass the zero-initialized residual map to supervise shared units. Retrospective evaluation spans four cohorts in two countries and multiple regions. With shared patient splits, labels, MRI preprocessing and scoring, baseline AUCs of 71.12% for I-SPY2 + Duke, 83.68% for Hospital-A and 79.56% for Hospital-B exceed those of retrained MRP–MRI and MORM in all three populations. Hospital-B AUC gains are 11.84% and 11.21%, respectively. Frozen expert restrictions support clinical–treatment complementarity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Treatment-response prediction, Patient-conditioned treatment modeling, Treatment-unit representation learning, Outcome-supervised representation learning, Longitudinal multimodal learning, Mixture-of-experts

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_X96DOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors have tackled a relevant real-life problem - treatment response prediction in multimodal settings face challenges of using varying historical context and diverse modalities which can be clinical, imaging or treatment related. They model concurrent and sequential relations, and evaluate on data across countries and cohorts.

Critical strengths
  1. The problem is practically motivated.

  2. Ablations across modalities are conducted.

  3. Results are reported on public dataset and private ones.

Critical weaknesses

The writing is hard to follow along, and does not provide adequate context. I would encourage the authors to consider simplifying the introduction to make the motivation and contributions more clear and accessible to the broader community. Broader comparisons against benchmarks like [1] would also be good to add.

Questions for authors
  1. In section 2.1, please add examples wherever possible to make the draft easier to read

  2. What does “unresolved regimens” mean in section 2.2?

  3. Why have separate experts instead of combining experts of individual inputs as needed?

  4. How does the proposed model compare against [1]? Is this infeasible for some reason?

  5. Table 1 shows a high variance across methods and dataset - is there a reason for this? Can it be alleviated?

  6. Table 2 shows that “without I” does quite well - is there any intuition for this?

[1] Tang, W., Jin, C., Kong, Q., Liu, C., Chen, S., Ding, S., ... & Guo, Y. (2025). Development and validation of an MRI spatiotemporal interaction model for early noninvasive prediction of neoadjuvant chemotherapy response in breast cancer: a multicentre study. EClinicalMedicine, 85.

Reviewer_a2zHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of predicting pathological complete response to neoadjuvant therapy while explicitly modeling patient-dependent interactions between treatment components and accommodating variable MRI histories and missing modalities. TheraRoute-L addresses the treatment-representation problem by preserving whole-regimen semantics while learning record-supported treatment units and symmetric concurrent or directed sequential relations that are queried using clinical and imaging context. A spatial–spectral MRI encoder and an observed-prefix history encoder represent patient imaging, while seven availability-masked modality-subset experts combine the available evidence.

Critical strengths
  1. The design is interesting and well motivated: the model preserves whole-regimen semantics while distinguishing documented units, concurrent combinations, and sequential relations. Retaining unresolved descriptions at their recorded granularity avoids unsupported decomposition and makes the approach suitable for heterogeneous treatment records.
  2. The design handles missing modalities. The seven private subset experts use only their declared inputs, and availability-masked routing combines eligible predictions without requiring modality imputation. Preserving the remaining experts’ raw outputs when a modality is removed provides a clearly defined prediction rule for incomplete records.
Critical weaknesses
  1. The two hospital datasets use different threshold-selection protocols: Hospital-A uses a fixed threshold of 0.5, whereas Hospital-B uses locally fitted thresholds. Both hospitals informed system development, raising concerns about potential development-time selection effects even though the final network parameters were frozen. Hospital-B’s classification thresholds use local labels through cross-fitting, so its threshold-based results reflect locally adapted performance rather than deployment with an entirely source-fixed decision rule.
  2. The ablation study is useful. However, relative to the whole-text-only variant A1, the full model A0 changes AUC by −0.19, +1.27, and +3.55 percentage points across the three populations; only Hospital-B’s paired confidence interval excludes zero. Moreover, the ablations change structured pooling and dedicated reads together, so they do not isolate the benefit of patient conditioning itself. The score-level mathematical results do not resolve this empirical attribution problem.
  3. Although the paper reports relevant ablations, they do not establish that patient conditioning itself is the main source of the performance improvement. A controlled comparison of patient-independent treatment reads and reads conditioned on clinical information only, imaging only, or both, with the remaining architecture and inputs unchanged, would materially strengthen the evidence.
Questions for authors
  1. Did hospital outcomes or evaluation results influence model design or tuning? Is there a sensitivity analysis that assesses this influence?
  2. Why does Hospital-A use a fixed threshold of 0.5 while Hospital-B uses locally fitted thresholds? Were these choices made before examining hospital performance?
  3. Could you compare patient-independent treatment reads with reads conditioned on clinical information only, imaging only, or both, while keeping the remaining architecture and training unchanged?
Reviewer_kDk5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

In this manuscript, the authors present TheraRoute-L, a model that predicts pathological complete response (pCR) to neoadjuvant therapy in breast cancer from breast MRI, clinical fields and treatment records. MRI examinations are encoded with a spatial–spectral 3D encoder over two contrast phases and a causal history transformer. Treatment text is parsed into units (named drugs, drug classes or unresolved regimens) and concurrent or sequential relations, encoded with frozen PubMedBERT, and scored by patient-conditioned low-rank bilinear maps. Seven modality-subset experts are combined by availability-masked routing. The authors also derive several elementary results (a three-way interaction bound, a Lipschitz bound on the treatment read, and gradient paths at zero initialization). On I-SPY2 + Duke (five-fold cross-validation) and two hospitals in a second country, TheraRoute-L reports AUCs of 71.1%, 83.7% and 79.6%, exceeding retrained MRP–MRI and MORM. Overall, the clinical problem is important and the multi-cohort evaluation is a strength, but the evidence does not support the central treatment-modeling contribution.

Critical strengths

Strengths

  1. The paper addresses an important clinical problem and explicitly models treatment composition and order, which prior multimodal response models usually reduce to coarse regimen categories.

  2. The evaluation spans four cohorts in two countries, applies frozen source models to the hospital cohorts, reports bootstrap confidence intervals and all six classification metrics, and is transparent about caveats such as outcome-associated treatment availability.

Critical weaknesses

Major issues

  1. The results do not support the paper's central contribution. The treatment-unit and relation ablations (A1/A2/A0) are described as yielding only small differences and are not reported in the main text. In Table 2, removing the image experts (C/T/CT) gives a higher AUC than the full model in all three populations, and the clinical-only expert nearly matches the full model (69.7/82.3/77.8% vs 71.1/83.7/79.6%), so most of the discrimination appears to come from routine clinical variables such as HER2, ER/PR and subtype. Yet the main comparators are an MRI-only branch of MRP and MORM. A simple clinical baseline (e.g., logistic regression or gradient boosting on the clinical fields plus regimen category) is essential. Treatment information is also available for only 35 of 189 scored Hospital-B records.

  2. The hospital cohorts are not independent external validation. Both hospitals “informed system development”, and the final design (L3) is selected from several historical systems compared on Hospital-B (Table 3); Hospital-B decision thresholds are also fit locally. The pretreatment availability of the clinical and treatment snapshots is unverified and treatment availability is outcome-associated (the full model reaches 90.9% AUC on treatment-eligible Hospital-B patients), which raises leakage concerns. The bootstrap intervals are conditional on fixed predictions and ignore training variance, and Hospital-A has only 96 patients. A hospital cohort untouched during development, or at least a precise account of which decisions each hospital informed, is needed.

Minor issues

  1. The theoretical results are elementary (a finite-difference identity for a bilinear form, a Lipschitz bound for a softmax-weighted tanh, and the chain rule) and are not connected to empirical behavior; they could move to the appendix to make room for the treatment ablation results.

  2. The paper is very dense and difficult to follow: many internal labels (L1–L3, V1–V5, A0–A2, P0/P1, IT/CT/ICT) are not defined in the main text, and essential details are deferred to appendices. The abstract's “baseline AUCs of 71.12%” is ambiguous, since “baseline” refers to the baseline MRI examination rather than a baseline model. The historical systems in Table 3 are the authors' own unpublished variants and add little to the comparison.

Questions for authors

N/A

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.