Ethical Fairness in Ubiquitous Health Sensing without Known Attributes

Shaily Roy; Harshit Sharma; Daniel A. Adler; Srijan Sen; Tanzeem Choudhury; Asif Salekin

cs.CYMar 10, 2026

Ethical Fairness in Ubiquitous Health Sensing without Known Attributes

Authors: Shaily Roy, Harshit Sharma, Daniel A. Adler, Srijan Sen, Tanzeem Choudhury, Asif Salekin

Organizations: Ira A. Fulton Schools of Engineering, Arizona State University, Tempe, Arizona, USA · Information Science, Cornell University, New York, New York, USA · University of Michigan, Ann Arbor, Michigan, USA · Cornell University, Ithaca, New York, USA

Abstract

In ubiquitous and mobile health systems, computational models infer human states from wearable, behavioral, and physiological sensing data. In these settings, high accuracy alone is insufficient; models must act ethically and equitably across diverse people, contexts, and devices. However, fairness methods that rely on demographic or heterogeneous attributes during training are difficult to enforce because such attributes are often unavailable, privacy-sensitive, regulated, or undesirable to collect. Conventional parity-based fairness can also violate ethical principles by trading off subgroup performance. To address this challenge, we present Flare, Fisher-guided LAtent-subgroup learning with do-no-harm REgularization, a demographic- and heterogeneous-attribute-agnostic framework that aligns human-centered fairness with ethical principles for ubiquitous and mobile sensing. Flare leverages optimization geometry, particularly Fisher Information, to regularize curvature and uncover latent disparities in model behavior without demographic or heterogeneous attributes. By integrating representation, loss, and curvature signals, it identifies hidden performance strata and refines them through collaborative but do-no-harm optimization, enhancing subgroup performance while preserving ethical balance. We also introduce BHE (Beneficence-Harm Avoidance-Equity), a metric suite that operationalizes ethical fairness beyond statistical parity. Across mobile physiological, behavioral, and clinical sensing datasets, including EDA, OhioT1DM, IHS, and Percept-R, Flare improves ethical fairness over state-of-the-art baselines. Ablation, interpretability, and loss-landscape analyses show that these gains arise from flatter optimization geometry, simpler decision rules, and do-no-harm latent-subgroup adaptation. Runtime analysis supports the practicality of Flare for resource-constrained sensing deployments.

Figures & tables

Figure 1. The Three-Stage Architecture for Flare . Stage 1 (Base PreTraining): The encoder-decoder framework generates latent embeddings $z$ and a pretrained model $\theta^{*}$ using a multi-task loss function. Stage 2 (Latent Clustering): Latent subgroups are derived from model behavior descriptors $s=[z,L_{CE},F]$ (embedding, cross-entropy loss, and Fisher), projected to a lower-dimensional space (e.g., via UMAP) and clustered (e.g., GMM) to obtain $C_{1},\dots,C_{n}$ clusters. Stage 3 (Cluster Adaptation and Aggregation): Cluster-specific individual models $\theta^{*}_{c}$ are initialized from $\theta^{*}$ ’s encoder and classifier layers, and are fine-tuned on CAA-train data for each cluster. At scheduled intervals, the aggregated model $\bar{\theta}=\frac{1}{n}\sum_{c=1}^{n}\theta^{*}_{c}$ is redistributed to all clusters as a candidate update. During this adaptation, a cluster adopts $\bar{\theta}$ only if it improves performance, otherwise its previous $\theta^{*}_{c}$ is retained for evaluation.

Dataset	Demographic attributes	Heterogeneous attributes
OhioT1DM ( 74 )	Age; Sex	Cohort; Pump; Sensor Band
EDA ( 118 )	Sex	Group_label
IHS ( 2 )	Sex; Age; Ethnicity	Specialty; PHQ10 $>$ 0
Percept-R ( 13 )	Sex; Age(months); Race; Ethnicity	Not available

Table 1. Sensitive attributes used during BHE evaluation.

Figure 2. K-fold person-disjoint evaluation overview. In each fold, participants are partitioned into a training pool and a fully unseen test set. The training pool is further divided into BpT-train and CAA-train subsets used for model development and selection. Final evaluation is performed on unseen person-disjoint test participants, and results are averaged across folds.

Sensitive Attribute	Baseline (%)	Fair-EO (%)	Fair-DP (%)	Flare (%)
OhioT1DM Dataset
Age	70.69 $\pm$ 2.79	69.16 $\pm$ 2.85	70.38 $\pm$ 5.09	74.36 $\pm$ 1.89
Cohort	68.12 $\pm$ 1.51	66.96 $\pm$ 2.52	68.41 $\pm$ 1.79	72.91 $\pm$ 0.66
Sex	68.25 $\pm$ 2.93	67.31 $\pm$ 1.94	68.67 $\pm$ 0.73	72.72 $\pm$ 2.05
Pump	69.97 $\pm$ 1.74	68.20 $\pm$ 0.66	68.13 $\pm$ 1.77	73.62 $\pm$ 0.64
Sensor Band	68.12 $\pm$ 1.51	66.96 $\pm$ 2.52	68.41 $\pm$ 1.79	72.91 $\pm$ 0.66

Table 2. F1 score (%) across sensitive attributes for different models in the OhioT1DM, EDA, IHS, and Percept-R datasets. For each sensitive attribute, F1 is computed separately for each subgroup, and the reported mean and standard deviation are calculated across the subgroup-level F1 scores of each sensitive attribute.

Sensitive Attribute	Fair-EO (%)			Fair-DP (%)			Flare (%)
Sensitive Attribute	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$
OhioT1DM Dataset
Age	$-1.53$	$-0.29$	$-0.06$	$-0.31$	$0.00$	$-2.31$	$\mathbf{3.67}$	$\mathbf{5.75}$	$\mathbf{0.90}$
Cohort	$-1.16$	$-2.79$	$-1.02$	$0.29$	$-0.83$	$-3.30$	$\mathbf{4.79}$	$\mathbf{4.47}$	$\mathbf{0.85}$
Sex	$-0.94$	$-1.38$	$0.99$	$0.42$	$0.84$	$\mathbf{2.20}$	$\mathbf{4.48}$	$\mathbf{3.96}$	$0.88$
Pump	$-1.77$	$-0.49$	$1.09$	$-1.84$	$-1.34$	$-0.02$	$\mathbf{3.66}$	$\mathbf{4.95}$	$\mathbf{1.11}$

Table 3. Comparison of the fairness-with-demographics models, Fair-EO and Fair-DP , against Flare . Values are expressed as percentage-point changes, with higher values indicating stronger ethical performance. The reported

\Delta B

\Delta H

, and

\Delta E

metrics are computed from subgroup-level performance statistics for each sensitive attribute

Dataset	Sensitive Attribute	Benign Baseline (%)	KD (%)	ARL (%)	Reckoner (%)	GOG (%)	Flare (%)
OhioT1DM	Age	70.69 $\pm$ 2.79	65.01 $\pm$ 2.44	70.16 $\pm$ 0.57	66.42 $\pm$ 2.21	59.12 $\pm$ 2.67	74.36 $\pm$ 1.89
	Cohort	68.12 $\pm$ 1.51	67.33 $\pm$ 0.37	70.46 $\pm$ 0.67	67.57 $\pm$ 0.96	63.50 $\pm$ 1.11	72.91 $\pm$ 0.66
	Sex	68.25 $\pm$ 2.93	69.08 $\pm$ 2.50	70.71 $\pm$ 0.20	68.51 $\pm$ 2.12	69.09 $\pm$ 2.64	72.72 $\pm$ 2.05
	Pump	69.97 $\pm$ 1.74	60.66 $\pm$ 2.81	65.80 $\pm$ 0.97	61.74 $\pm$ 1.23	68.01 $\pm$ 1.99	73.62 $\pm$ 0.64
	Sensor Band	68.12 $\pm$ 1.51	67.33 $\pm$ 0.78	70.46 $\pm$ 3.67	67.57 $\pm$ 0.96	65.00 $\pm$ 1.11	72.91 $\pm$ 0.66
EDA	Group_label	86.91 $\pm$ 4.47	69.20 $\pm$ 2.41	87.61 $\pm$ 3.34	85.55 $\pm$ 3.68	87.03 $\pm$ 4.58	89.08 $\pm$ 3.91

Table 4. F1 scores (%) across sensitive attributes for different models in the OhioT1DM, EDA, IHS, and Percept-R datasets. For each sensitive attribute, F1 is computed separately for each of its subgroups, and the reported values denote the mean

\pm

standard deviation across those subgroup-level F1 scores.

Dataset	Sensitive Attribute	ARL (%)			KD (%)			Reckoner (%)			GoG (%)			Flare (%)
	Sensitive Attribute	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$
OhioT1DM	Age	-0.54	-7.92	2.22	-5.68	-15.99	0.35	-4.28	-13.84	0.58	-11.58	-27.06	0.12	3.67	5.75	0.90
	Cohort	2.34	-0.10	0.84	-0.79	-5.44	-1.87	-0.54	-5.05	0.55	-4.62	-12.30	0.40	4.79	4.47	0.85
	Sex	2.46	3.25	2.73	0.84	1.00	0.43	0.26	-1.29	0.81	0.85	1.05	0.29	4.48	3.96	0.88
	Pump	-4.17	-9.26	0.78	-9.31	-17.33	-1.07	-8.23	-15.18	0.51	-1.95	-4.24	-0.25	3.66	4.95	1.11
	Sensor Band	2.34	-0.10	-2.16	-0.79	-5.44	0.73	-0.54	-5.05	0.55	-3.12	-12.30	0.40	4.79	4.47	0.85

Table 5. Model-wise percentage changes

(\Delta B,\Delta H,\Delta E)

for ARL, KD, Reckoner, GoG, and Flare across datasets. Values are reported as percentage-point improvements, where larger deltas indicate stronger fairness performance and greater ethical consistency across subgroups. The reported

\Delta B

\Delta H

, and

\Delta E

metrics are computed from subgroup-level performance statistics for each sensitive attribute, using subgroup-wise mean performance and subgroup-wise standard deviation

Figure 3. Fold and Cluster-wise F1 comparison across datasets within the proposed Flare framework. Bars show F1-scores for Base preTraining (BpT, light brown) as the intermediate Stage (1) and Flare (dark brown) as the final Stage (3) across all fold-cluster pairs. Values above each pair ( $\Delta$ ) indicate the change from BpT to Flare .

Figure 4. Overview of number of rules and BHE comparison across clusters C0–C3 in Fold 2. The top row shows the number of rules applied by the Benign Baseline, BpT, and FLARE models for SEX (Female, Male) and pump (530G, 630G) subgroups. The middle and bottom rows display percentage-point changes in $\Delta B$ , $\Delta H$ , and $\Delta E$ metrics for SEX and PUMP , respectively, relative to the Benign Baseline, highlighting differences between BpT and FLARE .

Individual (Sex-Pump)	True Label	Predicted Label	BpT: all fired rules	Flare : all fired rules	Cluster
Male-530G	0	BpT: 1 Flare: 0	Predicts 1 if: (1) 3hG $>0.64698$ ; w=1.000 . (2) Meal $\leq 0.4873$ & Work $>0.0984$ ; w=0.277 . (3) 3hG $>0.5657$ ; w=0.155 . (4) Work $>0.5738$ & 3hG $>0.5045$ ; w=0.118 . Predicts 0 if: (1) Bolus $\leq 0.6071$ & 3hG $>0.4354$ ; w=0.256. (2) TSB $\leq 0.2529$ , Bolus $>0.2322$ & Meal $\leq 0.5220$ ; w=0.253. (3) 3hG $>0.5311$ ; w=0.115. (4) Work $>0.9496$ ; w=0.056.	Predicts 0 if: (1) Meal $\leq 0.3791$ ; w=0.198 . (2) Work $>0.5952$ & Bolus $\leq 0.4174$ & Basal $\leq 0.7833$ ; w=0.148 . (3) Work $>0.4063$ & Meal $\leq 0.6927$ ; w=0.125 . Predicts 1 if: (1) Work $>0.3044$ & 3hG $>0.6195$ ; w=0.339.	0
Male-530G	1	BpT: 1 Flare: 1	Predicts 1 if: (1) 3hG $>0.625223$ & TSB $\leq 0.399254$ ; w=0.711 . (2) Meal $\in(0.373278,0.621851]$ & TSB $\leq 0.683481$ ; w=0.597 . (3) TSB $\leq 0.0558446$ ; w=0.3348 . (4) 3hG $>0.492165$ & Meal $>0.462781$ ; w=0.0317 . (5) Bolus $>0.347381$ & Basal $\leq 0.814583$ ; w=0.0213 . Predicts 0 if: (1) Meal $\in(0.481192,0.642121]$ ; w=0.340.	Predicts 1 if: (1) 3hG $>0.567255$ ; w=1.000 . Predicts 0 if: (1) Basal $\leq 0.722619$ ; w=0.50.	2

Table 6. Rule-based interpretation (Fold 2, OhioT1DM) for samples from a single individual (Male-530G) assigned to different clusters. For each model (BpT and Flare ), all fired rules are reported and grouped by predicted class . Multiple rules may be satisfied simultaneously , and the final prediction is obtained by aggregating the weights of all fired rules. Rules with the largest weights (dominant contributions) are bolded. Abbreviations: 3hG = three-hour glucose, TSB = time since last bolus.

Figure 5. Loss landscape visualizations for the benign baseline, base Pre-training without Fisher penalty, base pretraining with Fisher penalty regularization, and Flare for the OhioT1DM dataset. The red , orange , and green arrows and dotted circles respectively highlight curvature transitions, showing how Fisher regularization and cluster adaptation progressively flatten the optimization landscape.

Sensitive Attribute	n=1 (%)			n=2 (%)			n=3 (%)			n=4 (%)			Flare (%)
Sensitive Attribute	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$
Age	-9.25	0.50	-3.79	0.38	-1.58	-1.33	-1.48	-12.23	-7.90	5.56	3.17	0.71	3.67	5.75	0.90
Cohort	-4.64	-9.04	-7.93	4.44	1.09	-2.00	2.96	0.42	-1.19	6.09	3.65	-2.41	4.79	4.47	0.85
Sex	-3.85	-4.81	0.26	4.66	5.28	-0.11	3.15	3.61	2.46	5.01	3.15	-0.86	4.48	3.96	0.88
Pump	-14.82	1.84	-4.90	-0.10	-2.92	-2.82	-5.60	-13.57	-7.97	6.02	2.33	-0.31	3.66	4.95	1.11
Sensor Band	-4.64	-9.04	-4.92	4.44	1.09	-2.00	2.96	0.42	-1.19	6.09	3.65	-2.41	4.79	4.47	0.85

Table 7. Cluster sensitivity analysis showing percentage changes

(\Delta B,\Delta H,\Delta E)

for different numbers of clusters per fold (

K=1,2,3,4

) and Flare for OhioT1DM dataset. Values are reported in percentage points.

Dataset	Sensitive Attribute	BpT-w Fisher (%)			BpT-wo Fisher (%)			CAA-wo Fisher (%)			Flare (%)
	Sensitive Attribute	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$
OhioT1DM	Age	-9.25	0.50	-3.79	-1.24	-2.07	-1.82	1.86	-1.40	-3.78	3.67	5.75	0.90
	Cohort	-4.64	-9.04	-7.93	0.97	-0.43	-0.69	0.28	-0.49	-2.06	4.79	4.47	0.85
	Sex	-3.85	-4.81	0.26	0.68	-4.53	-12.37	-0.47	-2.06	-1.05	4.48	3.96	0.88
	Pump	-14.82	1.84	-4.90	-9.48	-3.41	-2.05	0.13	-0.74	0.83	3.66	4.95	1.11
	Sensor Band	-4.64	-9.04	-4.92	0.97	-0.43	-0.69	-0.28	2.49	-1.06	4.79	4.47	0.85

Table 8. Ablation study showing percentage changes

(\Delta B,\Delta H,\Delta E)

for Base-Training with Fisher (BpT-w Fisher), Base-Training without Fisher (BpT-wo Fisher), Cluster Adaptation and Aggregation without base-trained Fisher (CAA-wo Fisher), and Flare (CAA-w-Fisher) across datasets. Values are percentage points.

Table 9. Training and Inference runtime latency and resource comparison across deployment platforms for the EDA dataset (wall time in ms/sample; compute utilization in %; peak RSS memory in MB).

Phase	Flare pipeline stage	AMD CPU	NVIDIA GPU	Spark CPU	Spark GPU	Apple M4	Google Pixel 6	Raspberry Pi
Training	Pre-training stage	0.211	0.338	0.308	0.596	0.157	1.456	1.714
Training	Cluster fitting stage	5.416	5.630	6.365	5.769	11.472	10.815	17.691
Training	Training cluster assignment	0.012	0.013	0.014	0.013	0.015	0.014	0.012
Training	Adaptation for Cluster 1	0.382	0.733	0.554	0.904	0.299	3.471	3.093
Training	Adaptation for Cluster 2	0.348	0.672	0.542	0.739	0.245	3.449	3.126
Inference	Inference cluster assignment	2.360	2.387	2.870	2.881	2.023	4.476	10.338

Table 10. FLARE training and inference stage latency across platforms for the EDA dataset (wall time in ms/sample).

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Participants	Features	Samples	Task	Model (Encoder–Classifier)
OhioT1DM ( 74 )	11	7	8307	Glucose (binary)	input–256–128–64–32 / 32–8–2
IHS ( 2 )	85	18	2315	Mood (binary)	input–128–64–32 / 32–16–8–2
EDA ( 118 )	76	340	5497	Stress (binary)	input–128 / 128–64–64–2
Percept-R ( 13 )	76	(60,5)	32852	Speech correctness (binary)	input–52–32–32–2 / 52–32–2

Appendix

Table 11. Summary of datasets, total samples, and base model configurations.

	Stage 1			Step 3
	$lr$	$\alpha$	$\beta$	$lr$	$\alpha$	Frozen Layers
IHS	$3.28\times 10^{-4}$	0.497	0.0259	$5.09\times 10^{-4}$	0.568	4
Ohio	$1.0\times 10^{-3}$	0.74	0.919	$1.83\times 10^{-5}$	0.614	9
EDA	$2.30\times 10^{-3}$	0.76	0.79	$4.38\times 10^{-5}$	0.690	8
Percept-R	$1.0\times 10^{-3}$	0.3	0.7	$1.0\times 10^{-4}$	0.7	3

Appendix

Table 12. Hyperparameters used for Base Pretraining (Stage-1) and cluster-adaptation-aggregation (Stage-3).

	Percept-R			EDA
Metric	LSTM	MLP	CNN	LSTM	CNN	MLP
Accuracy	0.7639	0.7665	0.7789	0.6445	0.8529	0.8676
F1-score	0.7737	0.7665	0.7777	0.6418	0.8512	0.8692
Recall	0.7839	0.7865	0.7989	0.6445	0.8529	0.8676
AUC	0.7777	0.8485	0.8520	0.6617	0.9025	0.9223
	Ohio			IHS

Appendix

Table 13. Performance comparison across datasets. For each dataset, models are ordered from lowest to highest F1-score (best model on the right).

Sensitive Attribute	Subgroup	Baseline F1	KD F1	ARL F1	Reckoner F1	GOG F1	Flare F1
OhioT1DM Dataset
Sex	Female	0.6839	0.6985	0.7085	0.7136	0.6882	0.7127
	Male	0.6119	0.6631	0.7057	0.6937	0.6137	0.7417
Age	20-40	0.3942	0.5089	0.5896	0.6922	0.3983	0.7407
	40-60	0.6723	0.6976	0.7121	0.6931	0.6776	0.7263
	60-80	0.7103	0.7437	0.8030	0.7815	0.6976	0.7638

Appendix

Table 14. F1 scores across subgroups for different models (Baseline, KD, ARL, Reckoner, GOG, Flare ) in the OhioT1DM and EDA datasets.

Sensitive Attribute	Subgroup	Baseline F1	KD F1	ARL F1	Reckoner F1	GOG F1	Flare F1
Sex	Male	0.5975	0.5694	0.6203	0.5764	0.5857	0.6247
	Female	0.5957	0.5910	0.6324	0.5673	0.6128	0.6418
Age	23	0.8151	0.5884	0.8182	0.7273	0.4969	0.9091
	24	0.3194	0.2941	0.4158	0.3678	0.3529	0.4436
	25	0.5872	0.5772	0.5901	0.6101	0.5756	0.6033
	26	0.6515	0.6071	0.6588	0.5803	0.6635	0.6915

Appendix

Table 15. F1 scores across subgroups for different models (Baseline, KD, ARL, Reckoner, GOG, Flare ) in the IHS dataset.

Sensitive Attribute	Subgroup	Baseline F1	KD F1	ARL F1	Reckoner F1	GOG F1	Flare F1
Sex	Male	0.7638	0.7648	0.7611	0.7488	0.7639	0.7868
	Female	0.7817	0.7837	0.7637	0.7304	0.7890	0.8248
Age (months)	$\leq$ 100	0.7033	0.7445	0.6677	0.6515	0.7187	0.8007
	101-130	0.7574	0.7584	0.7523	0.7334	0.7545	0.7969
	131-160	0.8198	0.8005	0.8083	0.7851	0.8100	0.8141
	$>$ 160	0.8370	0.8360	0.8374	0.8379	0.8408	0.8593

Appendix

Table 16. F1 scores across subgroups for different models (Baseline, KD, ARL, Reckoner, GOG, Flare ) in the Percept-R dataset. Due to the high cardinality of age values in Percept-R, age is grouped into bins (chunks), and count-weighted aggregated F1 scores are reported per bin for readability and meaningful subgroup comparison.

Phase	Flare Pipeline Stage	AMD CPU		NVIDIA GPU		Spark CPU		Spark GPU		Apple M4		Google Pixel 6		Raspberry Pi
		Util.	RSS	Util.	RSS	Util.	RSS	Util.	RSS	Util.	RSS	Util.	RSS	Util.	RSS
Training	Pre-training stage	3.17	555.4	77.09	1022.4	4.99	488.4	7.56	1458.1	9.64	306.5	12.40	326.4	25.10	416.9
Training	Cluster fitting stage	3.13	995.0	42.97	1622.9	5.03	955.8	1.40	2069.8	4.64	716.3	12.41	605.1	24.44	762.3
Training	Training cluster assignment	0.00	982.5	40.00	1608.8	0.00	936.2	0.00	2060.4	0.07	716.3	0.00	512.3	0.00	747.1
Training	Adaptation for Cluster 1	2.56	1009.7	43.00	1648.6	4.84	956.7	0.00	2155.3	8.48	680.0	12.17	513.6	24.60	767.8
Training	Adaptation for Cluster 2	3.08	1009.8	84.38	1648.6	4.85	956.7	0.00	2155.3	9.29	680.6	12.34	515.4	24.90	769.3

Appendix

Table 17. FLARE training and inference stage latency and resource breakdown across deployment platforms for the EDA dataset (compute utilization in %; peak RSS memory in MB).

Table 18. Training latency and resource comparison across deployment platforms for the IHS, Ohio and PERCEPT-R datasets (wall time in ms/sample; compute utilization in %; peak RSS memory in MB).

Table 19. Inference latency and resource comparison across deployment platforms for the IHS, Ohio and PERCEPT-R datasets (wall time in ms/sample; compute utilization in %; peak RSS memory in MB).

Table 20. Flare training and inference stage latency and resource breakdown across deployment platforms for IHS, Ohio and PERCEPT-R datasets (wall time in ms/sample; compute utilization in %; peak RSS memory in MB)

Panel of Figure 6 — Figure 6. Loss landscape visualizations for the benign baseline, base Pre-training without Fisher penalty, base pretraining with Fisher penalty regularization, and Flare across EDA and IHS datasets. The red , orange , and green arrows and dotted circles respectively highlight curvature transitions, showing how Fisher regularization and cluster adaptation progressively flatten the optimization landscape.

Sex	Pump	True Label	BpT Rules	Flare Rules	Cluster
Female	630G	0	Predicts 0: fires when Meal $\in(-\infty,0.5152]$ , 3hG $\in(-\infty,0.52305]$ and Work $>0.2203$ ; weight 0.148. Predicts 0: fires when TSB $>0.09874$ and Bolus $>0.3947$ ; weight 0.249. Predicts 0: fires when 3hG $\leq 0.43715$ ; weight 0.096. Predicts 1: fires when Work $>0.8344$ and TSB $>0.06113$ and Meal $>0.3706$ ; weight 0.398 . Predicts 1: fires when Bolus $>0.32529$ and Meal $\in(-\infty,0.6124]$ and TSB $>0.2609$ ; weight 0.291 . Predicts 1: fires when Meal $\in(-\infty,0.4873]$ and Work $>0.09841$ ; weight 0.277 . Predicts 1: fires when Bolus $\in(0.6349,0.7972]$ ; weight 0.172 .	Predicts 0: fires when Work $\in(0.12936,\infty)$ and Bolus $>0.4174$ and Basal $\in(-\infty,0.7833]$ ; weight 0.262 . Predicts 0: fires when TSB $>0.23824$ ; weight 0.144 . Predicts 0: fires when Work $>0.40626$ and Meal $\in(-\infty,0.69275]$ ; weight 0.125 . Predicts 0: fires when Bolus $>0.49206$ ; weight 0.101 . Predicts 0: fires when 3hG $\leq 0.61948$ ; weight 0.092 . Predicts 1: fires when TSB $<0.159817$ ; weight 0.0645.	0
Male	530G	0	Predicts 0: fires when Bolus $\in(-\infty,0.60406]$ and TSB $\in(-\infty,0.31102]$ ; weight 0.082. Predicts 0: fires when 3hG $>0.51101$ ; weight 0.0018. Predicts 1: fires when TSB $\in(-\infty,0.05584]$ ; weight 0.3348 . Predicts 1: fires when 3hG $>0.52084$ and Work $>0.29589$ and Bolus $\in(-\infty,0.611998]$ ; weight 0.201 . Predicts 1: fires when Meal $>0.37059$ and Bolus $\in(-\infty,0.634921]$ ; weight 0.173 . Predicts 1: fires when 3hG $>0.49217$ and Meal $>0.46278$ ; weight 0.0317 . Predicts 1: fires when Meal $>0.5347$ ; weight 0.0152 .	Predicts 0: fires when Basal $\in(-\infty,0.72262]$ ; weight 0.50 .	1
Female	630G	1	Predicts 0: fires when Meal $\in(0.48119,0.64212]$ ; weight 0.940 . Predicts 1: fires when 3hG $>0.62522$ and TSB $\in(-\infty,0.39925]$ ;weight 0.511. Predicts 1: fires when Meal $\in(0.37328,0.62185]$ and TSB $\in(-\infty,0.68348]$ ; weight 0.097. Predicts 1: fires when TSB $\in(-\infty,0.05584]$ ; weight 0.035. Predicts 1: fires when 3hG $>0.49217$ and Meal $>0.46278$ ; weight 0.032. Predicts 1: fires when Bolus $>0.34738$ and Basal $\in(-\infty,0.81458]$ ; weight 0.021.	Predicts 0: fires when Basal $\in(-\infty,0.72262]$ ; weight 0.50. Predicts 1: fires when 3hG $>0.56726$ ; weight 1.000 .	1
Male	530G	0	Predicts 0: fires when Bolus $\leq 0.604$ and TSB $\leq 0.311$ ; weight 0.082. Predicts 0: fires when 3hG $>0.511$ ; weight 0.002. Predicts 1: fires when Meal $>0.885$ ; weight 0.238 . Predicts 1: fires when 3hG $>0.520$ , Work $>0.296$ , Bolus $\leq 0.612$ ; weight 0.201 . Predicts 1: fires when Meal $>0.371$ , Bolus $\leq 0.635$ ; weight 0.173 . Predicts 1: fires when 3hG $>0.492$ and Meal $>0.463$ ; weight 0.032 . Predicts 1: fires when Meal $>0.535$ ; weight 0.015 .	Predicts 0: fires when Basal $\leq 0.723$ ; weight 0.50 .	2
Female	530G	1	Predicts 0: fires when Basal $>0.526$ , TSB $>0.399$ , 3hG $>0.561$ ; weight 0.404 . Predicts 0: fires when TSB $>0.0598$ and Bolus $>0.254$ ; weight 0.374 . Predicts 0: fires when 3hG $>0.511$ ; weight 0.002 . Predicts 1: fires when 3hG $>0.492$ and Meal $>0.463$ ; weight 0.032. Predicts 1: fires when Meal $>0.535$ ; weight 0.015.	Predicts 0: fires when TSB $>0.221$ ; weight 1.0. Predicts 1: fires when TSB $>0.0579$ ; weight 1.0 . Predicts 1: fires when 3hG $>0.567$ ; weight 1.0 . Predicts 1: fires when Basal $>0.723$ ; weight 0.50 .	2

Appendix

Table 21. Qualitative rule-based interpretation of latent clusters with attribute annotations (Fold 2, OhioT1DM). Representative samples where Flare predicts correctly while BasePretrained predicts incorrectly, shown separately for clusters C0-C2. Abbreviations: 3hG = three-hour glucose, TSB = time since last bolus, Bolus = insulin bolus, Basal = basal insulin, Meal = meal intake, Work = work indicator. Bold weights highlight the rule group driving the model’s prediction.

Sensitive Attribute	n=1 (%)			n=2 (%)			n=3 (%)			n=4 (%)			Flare (%)
Sensitive Attribute	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$	$\Delta B$	$\Delta H$	$\Delta E$
EDA Dataset
Group_label	1.50	-0.51	2.13	1.78	0.67	-1.31	-0.22	1.48	-1.47	0.69	0.48	-0.01	2.16	0.70	0.57
Sex	1.49	-0.13	0.31	0.93	-0.90	-1.73	1.67	-0.62	-2.92	0.89	-0.70	-1.61	2.55	0.68	0.06
Mean	1.49	-0.32	1.22	1.35	-0.11	-1.52	0.72	0.43	-2.20	0.79	-0.11	-0.81	2.36	0.69	0.31
IHS Dataset

Appendix

Table 22. Cluster sensitivity analysis showing percentage changes

(\Delta B,\Delta H,\Delta E)

for different numbers of clusters per fold (

K=1,2,3,4

) and Flare across the EDA, IHS, and Percept-R datasets. Values are reported in percentage points.

Algorithmic Fairness

Algorithmic fairness studies how automated decisions distribute errors, opportunities, and harms across groups, and how to measure and reduce disparities.

See all 419 papers

Apr 17, 2026

Fairness Constraints in High-Dimensional Generalized Linear Models

Machine learning models often inherit biases from historical data, raising critical concerns about fairness and accountability. Conventional fairness interventions typically require access to sensitive attributes like gender or race, but privacy and legal restrictions frequently limit their use. To address this challenge, we propose a framework that infers sensitive attributes from auxiliary features and integrates fairness constraints into model training. Our approach mitigates bias while preserving predictive accuracy, offering a practical solution for fairness-aware learning. Empirical evaluations validate its effectiveness, contributing to the advancement of more equitable algorithmic decision-making.

Yixiao Lin, James Booth

aDepartment of Statistics and Data Science, Cornell University

Algorithmic Fairness

Ask PDF

Jun 22, 2026

FairBED: Incorporating Fairness into Bayesian Experimental Design

Bayesian experimental design (BED) has traditionally focused exclusively on collecting data that is maximally informative, with little consideration of other desirable dataset characteristics, such as fairness. We therefore extend BED to target data that is both informative and fair. To this end, we introduce FairBED, which provides a formalism for the fairness of datasets themselves by measuring how compatible they are with a fair world, in which the target quantity we wish to learn about is unrelated to the sensitive attributes we want to be fair to. We then use this to construct practical fairness-aware BED objectives that maximise expected information gain about the target variables, while penalising information about sensitive attributes and/or their coupling with the target. We further extend FairBED to active learning, and show empirically that models trained on data gathered using FairBED provide improved fairness-accuracy trade-offs compared to randomly acquired data and conventional BED.

Marcel Hedman, Emily Alger, Brieuc Lehmann +2

Department of Statistics, University of Oxford · Department of Statistical Science, University College London

Algorithmic Fairness Bayesian Experimental Design Dataset Bias

Ask PDF

Aug 16, 2025

FAIRVAR: Fair Federated Learning via Variance Regularization

Federated learning (FL) allows collaborative training of machine learning models across multiple parties without sharing raw data. However, heterogeneous data can cause some clients to have disproportionate influence on the global model, leading to disparities in their performance. Fairness, understood as reducing these disparities, is therefore a crucial concern in FL and has been addressed in various ways. We studied performance equitable fairness in FL, where the goal is to minimize performance disparities across clients. We evaluated several existing fairness-aware methods and introduce here a new gradient-variance-regularized method, implemented in two variants: FairGrad (approximate) and FairGrad* (exact). We theoretically characterize the connections between these methods and, empirically, on heterogeneous benchmarks, show that FairGrad and FairGrad* consistently improve fairness by reducing variance in client accuracies, while maintaining competitive or improved mean performance compared to existing fairness-aware baselines.

Zahra Kharaghani, Ali Dadras, Tommy Löfstedt

Department of Computing Science, Umeå University, Umeå, Sweden · Department of Mathematics and Mathematical Statistics, Umeå University, Umeå, Sweden

Algorithmic Fairness Federated Learning Heterogeneous Federated Learning

Ask PDF

Abstract

Algorithmic Fairness

Explore similar work

Fairness Constraints in High-Dimensional Generalized Linear Models

FairBED: Incorporating Fairness into Bayesian Experimental Design

FAIRVAR: Fair Federated Learning via Variance Regularization