MO-OBAM: Multi-Objective Data Anonymization

• MO-OBAM is a multi-objective optimization-based anonymization framework designed for structured healthcare data, balancing privacy risk reduction with predictive utility.
• It employs clustering and microaggregation techniques to minimize both linkage and homogeneity risks, achieving up to 90% reduction in vulnerable records compared to other methods.
• The framework maintains high data utility with only about a 2% change in precision and recall, outperforming basic k-anonymity and other optimization-based techniques.

Searching arXiv for the MO-OBAM paper and closely related anonymization work to ground the article. MO-OBAM denotes the Multi-Objective Optimization-Based Anonymization Model, an optimization-based data anonymization framework for structured, tabular datasets intended for downstream machine learning, with a particular emphasis on healthcare data sharing. In the study that explicitly defines the term, MO-OBAM is evaluated on retrospective EHR data for sepsis prediction and is presented as a method that simultaneously reduces privacy risks—especially linkage attacks and homogeneity attacks—and preserves predictive utility for downstream classifiers. Within that evaluation, MO-OBAM provides stronger privacy safeguards than a basic $k$-anonymity microaggregation algorithm and an optimization-based $l$-diversity method by Zheng et al., while yielding the best utility outcomes, with only about a 2% change in precision and recall relative to the original dataset (Wei et al., 25 Aug 2025).

1. Definition and problem formulation

MO-OBAM is situated in the privacy-preserving data publishing literature as a model-based anonymization approach designed for settings in which a healthcare organization must share EHR-derived data with ML practitioners while protecting patient privacy. Its purpose is to simultaneously reduce privacy risks—especially for vulnerable sub-populations—and preserve as much predictive performance as possible for a specified downstream ML task, here binary classification of sepsis (Wei et al., 25 Aug 2025).

The underlying application uses retrospective EHR data from a tertiary-care system covering adult patients admitted between July 2013 and December 2015. The dataset contains 119,871 unique individuals and 106 variables per patient. Among these variables, the study defines 6 quasi-identifiers (QIs)—Age, Length of stay, Number of visits during the study period, Gender, Race, and Ethnicity—together with 30 sensitive attributes (SAs) and 70 non-sensitive attributes (NSAs). The target variable is SepsisFlag, a binary indicator defined from sepsis-related ICD codes (Wei et al., 25 Aug 2025).

The paper frames anonymization decisions around three questions: whether to further anonymize beyond HIPAA, whether to include vulnerable individuals in the shared dataset, and which anonymization method best balances privacy and utility. MO-OBAM is proposed for the third of these questions and is described as enabling the sharing of anonymized data that still includes vulnerable individuals while significantly reducing their risk and keeping ML performance nearly unchanged (Wei et al., 25 Aug 2025).

2. Privacy risks and vulnerable populations

The privacy analysis focuses on two classical attacks. A linkage attack occurs when an attacker matches records in the shared dataset to external data using QIs; the risk is higher when QI combinations are rare or unique. A homogeneity attack occurs when all records in an equivalence class share the same sensitive attribute value, allowing an attacker to infer that value for every record in the class (Wei et al., 25 Aug 2025).

For linkage risk, records are grouped by exact QI combinations or their generalized versions. If a record belongs to a cluster $C_i$ of size $|C_i|$, its linkage risk is defined as

$$\text{Risk}_\text{linkage}(x \in C_i) = \frac{1}{|C_i|}.$$

A record is treated as vulnerable when this value exceeds a threshold $\tau \in \{0.05, 0.075, 0.1\}$. For homogeneity attacks, a record is vulnerable when all records in its cluster have the same value for a given sensitive attribute (Wei et al., 25 Aug 2025).

On the original HIPAA-compliant but not further anonymized data, the study reports the following baseline vulnerability counts:

• Linkage risk:
  • $\tau = 0.05$: 56,113
  • $\tau = 0.075$: 48,396
  • $\tau = 0.1$: 41,445
• Homogeneity attack risk: 14,937 individuals (Wei et al., 25 Aug 2025)

The paper also identifies disproportionately vulnerable sub-populations: older adults, those with long length of stay, patients with a high number of visits, males, Black or African American individuals, and Hispanic/Latino groups. These populations are not only privacy-sensitive but also clinically important for the downstream sepsis task. The study reports that vulnerable individuals are more likely to have sepsis than the overall population; for example, under linkage risk with $\tau=0.1$, the overall sepsis prevalence is 16.67%, whereas the vulnerable population has 20.85%. Under homogeneity-based vulnerability, the vulnerable group has 25.40% sepsis versus 16.67% overall (Wei et al., 25 Aug 2025).

This coupling between privacy vulnerability and predictive relevance is central to the rationale for MO-OBAM. The paper argues that excluding vulnerable individuals may reduce re-identification risk, but it also removes records that are disproportionately informative for model learning.

3. Anonymization design and optimization perspective

MO-OBAM is described as an optimization model operating on structured tabular data with mixed numeric and categorical QIs. It uses clustering of records based on QIs and then applies microaggregation, replacing original QI values by cluster-level aggregates. The paper states that the detailed mathematical formulation is given in Wei et al. (2025), but it provides the operational structure used in the study (Wei et al., 25 Aug 2025).

For all three anonymization methods evaluated in the study—basic $l$0-anonymity, Zheng et al.’s optimization-based $l$1-diversity method, and MO-OBAM—the actual transformation of QIs is identical:

• For numeric QIs—Age, Length of stay, and Number of visits—each original value is replaced with the mean of that QI within its cluster.
• For categorical QIs—Gender, Race, and Ethnicity—each original value is replaced with the mode within its cluster (Wei et al., 25 Aug 2025).

The difference between the methods lies not in the aggregation step, but in how clusters are formed, how large they are, and what constraints or objectives govern them. MO-OBAM is distinguished by treating anonymization as a multi-objective optimization problem. The paper characterizes its objectives as minimizing linkage risk, minimizing homogeneity risk, and minimizing utility loss relative to the original dataset. In the notation given in the study, this perspective is expressed as minimizing

$l$2

$l$3

and

$l$4

subject to clustering and diversity constraints (Wei et al., 25 Aug 2025).

The paper further states that, in practice, a weighted-sum or Pareto solution is chosen, although it does not provide the specific weights. This suggests that MO-OBAM is intended not merely to satisfy anonymity constraints mechanically, but to select anonymizations that remain useful for a predefined predictive task.

4. Experimental workflow and empirical results

The experimental pipeline begins with baseline risk evaluation on the original data, followed by anonymization at protection levels $l$5, re-evaluation of linkage and homogeneity risk on the further anonymized datasets, and then ML training on several candidate datasets. The downstream classification task is sepsis prediction. To reduce class imbalance for model evaluation, the study applies propensity score matching, producing 19,255 controls and 3,851 sepsis cases. The classifiers evaluated are Decision Tree, Logistic Regression, Gaussian Naive Bayes, Neural Network, Random Forest, and Support Vector Machine, each trained and tested 100 times with 80/20 train/test splits (Wei et al., 25 Aug 2025).

The privacy results show that MO-OBAM reduces vulnerable records more aggressively than the alternatives at the same protection level. For $l$6, the reported vulnerability counts are as follows:

Method	Linkage risk at $l$7	Homogeneity attack risk
Original data	41,445	14,937
$l$8-anonymity	35,480	6,905
Zheng et al.	7,486	0
MO-OBAM	2,954	0

At higher protection levels, MO-OBAM further reduces risk. For $l$9, it yields 6,939 vulnerable records at $C_i$0, 1,960 at $C_i$1, and 0 at $C_i$2, with 0 homogeneity vulnerabilities. For $C_i$3, MO-OBAM yields 2,020 vulnerable records at $C_i$4 and 0 at the two stricter thresholds, again with 0 homogeneity risk. For $C_i$5, all reported risks are 0 (Wei et al., 25 Aug 2025).

The study also reports subgroup effects. At $C_i$6 and $C_i$7, basic $C_i$8-anonymity reduces vulnerability by less than 20% across most QI categories, Zheng et al.’s method by roughly 70–80%, and MO-OBAM by roughly 90% across most QIs. This pattern is reported for age groups, length-of-stay ranges, visit-count categories, gender, race, and ethnicity (Wei et al., 25 Aug 2025).

5. Utility preservation and comparison with alternative anonymization methods

The paper emphasizes that privacy protection cannot be evaluated independently of downstream utility. It first compares the original matched dataset with a version containing only non-vulnerable individuals. The result is a substantial degradation in predictive performance. For example, the reported precision drops for OR vs OR-NV-only include -27.36% for Decision Tree, -18.69% for Logistic Regression, -42.36% for Naive Bayes, -15.17% for Neural Network, -12.18% for Random Forest, and -50.53% for SVM at $C_i$9. Reported recall drops are similarly large, including -26.69% for Decision Tree, -41.52% for Logistic Regression, -35.00% for Neural Network, -42\%$|C_i|$0-59\% for Random Forest depending on risk setting, and -34\%$|C_i|$1-44\% for SVM (Wei et al., 25 Aug 2025).

Against that baseline, anonymization with retention of the full population performs much better. For OR vs FA, the paper reports that precision changes under MO-OBAM are typically in [-1%, +4%], and recall changes are around [-0.8%, +1%] for most models, with some larger deviations for SVM. The study summarizes this behavior by stating that MO-OBAM yields only a 2% change in precision and recall compared to the original dataset (Wei et al., 25 Aug 2025).

The comparative positioning of the three anonymization methods can be summarized as follows:

Method	Privacy characteristics	Utility characteristics
$|C_i|$2-anonymity	Modest linkage-risk reduction; homogeneity attacks not fully mitigated	Moderate performance preservation
Zheng et al.	Strong privacy; homogeneity attack risk reduced to zero quickly	Performance close to original, but often slightly worse than MO-OBAM
MO-OBAM	Strongest linkage-risk reduction; homogeneity vulnerabilities eliminated	Best utility outcomes; about 2% change in precision and recall

The study also compares FA vs FA-NV-only. Under this comparison, removing vulnerable individuals from already anonymized data still reduces performance, but much less under MO-OBAM than under the alternatives. Reported MO-OBAM precision decreases are only about 0.3–1.4%, and recall decreases are typically 0.2–3%, often around 1–2%. For comparison, the corresponding performance losses are larger under Zheng et al.’s method and markedly larger under basic $|C_i|$3-anonymity (Wei et al., 25 Aug 2025).

A further reported result is OR-NV-only vs FA-NV-only: applying anonymization to the non-vulnerable-only datasets actually improves precision and recall. The paper reports average improvements under MO-OBAM of about 20–22% in precision and 29–39% in recall, larger than those observed for the other two methods (Wei et al., 25 Aug 2025).

6. Interpretation, implications, and limitations

The main interpretation advanced in the study is that vulnerable sub-populations should not be removed from shared healthcare data simply because they are at higher privacy risk. Those same groups are often clinically important and disproportionately associated with the target outcome. In the sepsis use case, older adults, patients with high utilization, and racial or ethnic minorities are both more vulnerable to re-identification and more important for predictive performance. The paper therefore recommends including vulnerable populations in the shared anonymized dataset and using an advanced optimization-based method such as MO-OBAM rather than basic $|C_i|$4-anonymity (Wei et al., 25 Aug 2025).

In broader methodological context, the paper places MO-OBAM alongside classical syntactic privacy models such as $|C_i|$5-anonymity, $|C_i|$6-diversity, and $|C_i|$7-closeness, as well as optimization-based anonymization approaches by Doka et al., Liang and Samavi, Aminifar et al., Zheng et al., and Wei et al. It presents MO-OBAM as especially suitable when there is a specific, well-defined ML task, when both privacy and performance objectives must be satisfied, and when the data contain both numeric and categorical QIs (Wei et al., 25 Aug 2025).

The reported limitations are also explicit. The study is based on a single dataset and outcome; SepsisFlag is derived from ICD codes used for billing; the evaluation considers only linkage and homogeneity attacks; and the detailed optimization formulation, solver choice, and computational complexity are not reproduced in the paper but deferred to Wei et al. (2025). The method is evaluated on labeled data, and the paper notes that applications to unlabeled settings may require different utility proxies (Wei et al., 25 Aug 2025).

Taken together, these results define MO-OBAM as a task-aware, multi-objective anonymization framework for healthcare data sharing: one that operates through cluster formation and mean/mode microaggregation of QIs, explicitly balances privacy and ML utility, sharply reduces vulnerable-record counts under linkage and homogeneity risk models, and preserves downstream predictive performance far better than exclusion-based strategies or simple $|C_i|$8-anonymity (Wei et al., 25 Aug 2025).