AI Reads Sinuses and Nose to Identify Sex from Skeletons

When forensic investigators recover human remains that are fragmented, burned, or decomposed, one of the first questions they must answer is deceptively simple: was this person male or female? The skull and pelvis usually provide the answer, but those bones are often missing or damaged. Now a team of Egyptian researchers has shown that two structures deep inside the face—the maxillary sinuses and the nasal cavity—can carry enough information to make that call, especially when machine learning is brought in to read the measurements. The study, published in the International Journal of Legal Medicine, offers a cautiously optimistic assessment of how far artificial intelligence can push the limits of forensic identification.

The maxillary sinuses are air-filled cavities hollowed out of the cheekbones, one on each side of the nose. Because they sit within some of the densest bone of the midface, they frequently survive fires, explosions, and other events that destroy more fragile skeletal landmarks. Forensic scientists have long suspected that these cavities differ between men and women, since male skulls tend to be larger and more robust overall. Earlier studies using cone-beam computed tomography in Egyptian, Indian, Chinese, and Yemeni populations confirmed measurable differences in sinus volume and dimensions. What remained unclear was how reliably those differences could be combined into a single, defensible classification—and whether the nasal structures nearby could add discriminating power.

To find out, researchers led by Haidy M. Fakher of Benha University and Eman S. Shaltout of Assiut University assembled a comparative cross-sectional dataset of 195 adult Egyptians: 100 females and 95 males, drawn from two geographically distinct regions. One hundred four participants came from Assiut in Upper Egypt, and 91 from Benha in Lower Egypt. Each person had undergone a computed tomography scan of the paranasal sinuses, the standard CT protocol used to image this region. Written informed consent was obtained from all participants, and the study received approval from the Assiut University Committee on Research Ethics, in compliance with the Declaration of Helsinki.

From each scan, the team extracted six dimensions of the maxillary sinuses and three nasal measurements, including the nasofrontal angle and the distance from the nasion—the bridge of the nose where it meets the forehead—to the nasal tip. They also recorded age and engineered ten additional derived features, bringing the total to 20 variables for the machine learning analysis. This feature engineering step matters: raw anatomical distances often carry redundant or correlated information, and ratios or combined indices can reveal dimorphism that individual measurements miss. The researchers then faced a classic machine learning design question: in what order should feature selection and hyperparameter tuning be performed?

They answered it by building two competing analytical frameworks. Framework 1 and Framework 2 differed in the sequence in which feature selection and hyperparameter optimization were applied. Within each framework, the team tested six classifiers—among them Random Forest, Support Vector Machine, and gradient boosting methods such as XGBoost, LightGBM, and CatBoost—crossed with five feature-selection methods. Performance was evaluated with a battery of standard metrics: accuracy, the area under the receiver operating characteristic curve (AUC), precision, recall, F1-score, and specificity. This kind of systematic sweep is increasingly common in forensic anthropology, where researchers have moved from traditional discriminant function analysis toward algorithms capable of capturing nonlinear relationships among skeletal measurements.

The results showed statistically significant sex-related differences in several of the measurements, though with notable regional variation between Upper and Lower Egypt—a reminder that craniofacial morphology is shaped by both sex and population history. Framework 2, which applied feature selection and hyperparameter optimization in a different order, generally outperformed Framework 1. The best-performing models achieved AUCs of 0.771 and 0.768, and the highest accuracy reached 74.4 percent. Those numbers describe what statisticians call moderate discriminative power: far better than a coin flip, and genuinely useful as corroborative evidence, but not yet in the 90-plus percent range achieved when the full skull or pelvis is available.

Perhaps the most practically valuable finding was which features carried the most weight. The nasofrontal angle, the nasion-to-tip distance, and the mean anteroposterior maxillary dimension emerged as the most consistent predictors across models. In plain terms, the shape of the nose where it meets the brow and the front-to-back depth of the cheek sinus did much of the classification work. That nasal measurements proved robust is significant, because the external nose itself is cartilage and rarely survives, but its bony framework—the nasal aperture, the nasion, the underlying cavity architecture—is preserved in bone and easily captured on a CT scan of the head.

The study’s regional findings add an important caveat for practitioners. Because the strength of sexual dimorphism varied between the Assiut and Benha samples, models trained on one population may not transfer cleanly to another. This echoes a well-established principle in forensic anthropology: sex estimation standards are population-specific, and applying a formula derived from one ancestry group to remains from another can introduce systematic error. The Egyptian data suggest that even within a single country, north–south geographic variation is large enough to register in sinus and nasal metrics, which argues for building locally calibrated models rather than importing thresholds from other populations.

The methodological comparison embedded in the study also carries lessons for the growing field of machine learning in forensic anthropology. Recent reviews have catalogued a rapid expansion of AI applications in this discipline, from cranial data mining to deep learning on cephalometric images, but results vary widely depending on pipeline choices. By showing that the ordering of feature selection and hyperparameter optimization measurably affects performance, the Egyptian team highlights a decision point that many published studies leave unexamined. Their use of interpretable model-agnostic importance analysis, rather than treating the classifiers as black boxes, also aligns with a broader push for transparency in algorithms that may one day inform court testimony.

Where does this leave the sinuses in the forensic toolkit? The authors conclude that CT-based nasal and maxillary sinus anthropometry shows moderate utility for sex estimation in Egyptians, with machine learning improving classification by capturing complex morphometric relationships that simpler statistical methods miss. For identification teams working with incomplete remains, the message is that these protected midface structures can serve as a meaningful secondary indicator when primary skeletal landmarks are absent—best used alongside other evidence rather than in isolation. With no external funding and data available from the corresponding author upon reasonable request, the study stands as a careful, population-anchored step toward AI-assisted identification, and a demonstration that even the hidden air spaces of the face can testify when the rest of the skeleton cannot.

Subject of Research: Machine learning-based forensic sex estimation from CT-derived nasal and maxillary sinus measurements in an Egyptian population

Article Title: Machine learning-driven forensic sex prediction using CT-based nasal and maxillary sinus metrics

Article References: Machine learning-driven forensic sex prediction using CT-based nasal and maxillary sinus metrics. (n.d.). https://doi.org/10.1007/s00414-026-03951-6

Image Credits: AI Generated

DOI: 10.1007/s00414-026-03951-6

Keywords: forensic anthropology, sex determination, machine learning, computed tomography, maxillary sinus, nasal morphometry, sexual dimorphism, Egyptian population, feature selection, Random Forest, support vector machine, International Journal of Legal Medicine