Beyond Demographic Balance: Multi-Metric and Intersectional Evaluation of Fairness in MIMIC-IV Mortality Prediction

The methodological clarity around isolating demographic representation from outcome-conditioned balancing is a genuine strength. By explicitly excluding the mortality label from the balancing rule, the authors enable a focused causal interpretation of what the intervention actually changes: "The label $y$ does not enter the grouping criterion. The intervention therefore enforces equal sample counts across observed demographic intersections, $|D^{\mathrm{bal}}_c|=n_{\min}\quad\forall c\in\mathcal{C}$, but does not impose a target value for $P(Y=1\mid(E,G,I)=c)$ or directly constrain $\mathrm{TPR}_c$ or $\mathrm{FPR}_c$." Additionally, the authors handle statistical fragility responsibly, foregrounding better-supported subgroups and reporting confidence intervals rather than treating sparse intersectional point estimates as authoritative: "The main text focuses on better-supported intersections with $N\geq 100$ and $N_{+}\geq 5$; Appendix B reports all intersections with $N_{+}\geq 5$, together with 95% Wilson intervals for sensitivity."

The primary limitation is data sparsity at the intersectional level, which constrains the reliability of fine-grained claims even after filtering. The authors note that "Intersection sizes range from 9 to 817 patients (median 45.5), with 1–101 mortality-positive cases per intersection (median 5)," so some quoted subgroup sensitivities hinge on very few events and wide Wilson intervals. A second limitation is that the intervention itself is minimal—pure downsampling followed by last-layer retraining—and is not benchmarked against outcome-conditioned balancing or explicit regularizers, extensions the authors identify as necessary future work: "A useful extension would be to compare demographic-only balancing directly with demographic–outcome balancing and with variants that incorporate an explicit fairness regularizer at the intersectional subgroup level."

The empirical evidence directly supports the two research questions. Figure 2 and Figure 3 convincingly demonstrate that reweighting achieves low FPR only by collapsing TPR, as "the median TPR is only approximately 0.15–0.45, compared with roughly 0.70–0.96 for the baseline and other interventions," while IBHFT lowers FPR without sacrificing sensitivity as severely. The intersectional audit provides concrete evidence that marginal summaries mislead: "For Black patients, the marginal sensitivity is 0.805 ... whereas Black female patients in the Other insurance category have sensitivity 0.400." Comparisons to related work in Table 1 are fair and transparently position the study alongside Mao et al., Lett et al., and Yoon and Kwak without overstating novelty.

Independent reproduction is feasible in principle but blocked by several missing details. The authors use the public MIMIC-IV dataset and specify base-model optimizer settings—"The reported configuration uses Adam with learning rate $10^{-3}$, batch size 200, and a maximum of 20 epochs"—and provide pseudocode for IBHFT in Algorithm 1, including "Freeze the base network through its 128-dimensional representation. Replace the original output layer with an MLP head $128\!\rightarrow\!512\!\rightarrow\!256\!\rightarrow\!1$." However, no code is released, the number of repeated evaluations is not stated, and random seeds and validation splits are omitted, so reproducing the exact variance estimates and subgroup rankings would require substantial guesswork.