BFHD: Bidirectional Feature Harmonization Decomposition for Heterogeneous Clinical Assessments
Clinical assessments collected across heterogeneous systems create mixed longitudinal records that resist standard total-score linking. This paper proposes Bidirectional Feature Harmonization Decomposition (BFHD), a method that identifies which individual measurements can be reliably translated between two assessment systems in both directions while explicitly separating out non-translatable components. By operating directly in original measurement spaces via a shared invertible backbone and feature-wise output gating, BFHD aims to provide an interpretable decomposition for cross-protocol integration and longitudinal analysis.
BFHD presents a conceptually elegant formulation of measurement-level harmonization as a bidirectional recoverability problem, and its neural instantiation is technically coherent. The paper is well-structured, the synthetic validation is convincing, and the ablation studies confirm that the proposed components (gate reward, hinge losses, shared backbone) are necessary. However, the central limitation is that real-world evaluation is self-referential: without ground-truth recoverability labels, the paper measures coverage using its own feasibility criterion $\varepsilon_{\text{cutoff}}$, and several pragmatic choices---zero-padding to align dimensions and mean imputation under missingness---are clinically naive. The work is a solid conceptual advance but needs external clinical validation to support deployment claims.
The problem formulation is principled and well-motivated. Defining recoverability relative to a bidirectional model class with a feasibility tolerance $\varepsilon$ is a sensible extension of classical linking. The hard-concrete gating mechanism together with the hinge-style losses $\mathcal{L}_{X \leftarrow Y} = \sum_{i=1}^{d_X} G_{X,i} \, \mathrm{ReLU}\bigl(\mathrm{err}(X_i, \tilde{X}_i) - \varepsilon_{\text{train}}\bigr)$ translates the combinatorial objective into a tractable differentiable proxy, and the ablation study confirms that replacing the gate reward with a standard $L_0$ penalty collapses to degenerate solutions. On synthetic data with known shared/private structure, BFHD recovers ground-truth labels accurately ($\mathrm{Acc}_X = 0.880$, $\mathrm{Acc}_Y = 0.830$) with high Jaccard stability.
The evaluation on real data is circular: baseline recoverability is inferred post hoc using the same $\varepsilon_{\text{cutoff}}$ feasibility criterion that BFHD optimizes, so claims of broader coverage are partially tautological. Additionally, the NICE-style invertible backbone requires zero-padding to $D = \max(d_X, d_Y)$, which introduces artificial dimensions that pass through the network despite being "never supervised"---a pragmatic hack that undermines the interpretability of the decomposition.
The temporal deployment results (100% retention with zero variance on EMBIC$\leftrightarrow$NeuroBAT) are suspiciously clean and involve tiny recoverable subsets, raising questions about cohort homogeneity rather than method robustness. Appendix C.2 further reveals clinically naive modeling: test-time missingness is handled via mean imputation, a crude expedient that can bias retained feasibility sets.
The evidence strongly supports BFHD on synthetic data because the ground-truth factor model provides independent labels for recoverable features. On real clinical pairs, the evidence is more mixed. Comparisons to predictive baselines (XGB, FRF) and IRT are methodologically mismatched: these methods are not designed for feature-level bidirectional decomposition, so outperforming them on $\#\mathrm{Rec}$ under $\varepsilon_{\text{cutoff}}$ is expected. Comparisons to CMLP and NICE are fairer, and BFHD shows modest improvements in Jaccard stability and coverage. The downstream prediction experiment (Appendix C.1) reveals that translated measurements underperform observed measurements for all methods, with BFHD showing the smallest accuracy drop---a useful but incremental gain.
The authors provide code and report detailed hyperparameters, including the NICE backbone widths $(64,64,32,32)$, learning-rate schedules, and the staged training procedure. The 10-seed stability analysis is welcome. However, independent reproduction is hindered by the access-controlled ADHD dataset and the use of greedy search for backbone selection without documenting the search space or initialization protocol. The three-stage training pipeline---with annealed $\varepsilon_{\text{train}}$, separated learning rates for backbone and gates, and gate hardening before fine-tuning---is complex and sensitive; small deviations could destabilize the gating behavior. Moreover, because gates are binarized after training, the exact hardening protocol and its interaction with the hard-concrete temperature $(1/8)$ are critical and only briefly described.
Clinical assessments are often collected using heterogeneous assessment systems across centers and time, leading to records that mix different but related sets of measurements. This motivates harmonization beyond total-score linking. We formulate clinical harmonization at the measurement level as a bidirectional recoverability problem: given paired observations from two assessment systems, the goal is to identify which measurements can be reliably translated in both directions within an application-defined tolerance, while separating non-translatable components. We propose Bidirectional Feature Harmonization Decomposition (BFHD), a feasibility-driven framework that enforces bidirectionally coupled translation and uses feature-wise output gating to produce an explicit decomposition in the original measurement space. Experiments on synthetic data and real clinical assessment pairs show that BFHD achieves broader feasible harmonization coverage and improved subset stability compared to baselines.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.