What we claim,
we publish.

What the study asked

Implants shaped to an individual patient's anatomy are now routine in cranial, facial, jaw and complex orthopaedic reconstruction. They are designed through a pipeline with three technical pillars: imaging, computational planning and, increasingly, artificial intelligence.

Earlier reviews asked whether a patient-specific implant performs as well as a conventional one. This review asked a different question: what evidence supports each pillar of the design pipeline itself, at which tier of evidence, and how credible that evidence is.

How it was done

The protocol was committed to a version-controlled repository before any record was retrieved, and the review is reported under PRISMA 2020, PRISMA-S and SWiM. MEDLINE, Europe PMC, Crossref and OpenAlex were searched on 2 and 3 September 2026 for primary studies from 2015 onwards. Engineering indexes were not available, so the review maps the biomedical-indexed literature and measures separately how much it misses.

Screening, charting and risk-of-bias appraisal were performed by AI systems from two different vendors, orchestrated by the authors under the protocol. Every record was screened by two independent screeners, every clinical study was appraised twice, and a random sample of 60 records was charted a second time. The paper states plainly that no human-verified subsample exists and reports how far the machine readings agreed with each other.

What it found

• Of 5059 records, 639 studies were included: 373 technical studies, 179 clinical series, 75 comparative studies and 12 randomised trials. Computational planning was applied in 601, an evaluated imaging step in 219 and artificial intelligence in 36.
• Randomised evidence came from 12 small trials, 10 of them craniomaxillofacial, and they evaluated the delivered implant rather than the planning method. No trial was at low risk of bias, and 69 of 75 comparative studies were at serious risk, mostly from unadjusted confounding.
• Pooling mostly non-randomised comparisons, patient-specific implants deviated less from the virtual plan (four studies; standardised mean difference −1.21, 95% CI −1.69 to −0.74) and were revised or removed less often (eight comparative studies, none randomised; risk ratio 0.28, 95% CI 0.11 to 0.72). All estimates are of very low certainty.
• Among technical studies read in full, independent validation material was used in fewer than half, and files were shared in about one in fourteen.
• Of 36 artificial intelligence studies, 4 were evaluated on independent data, 2 were used for a real patient and none were clinically comparative. The review received no external funding.

Why it matters

It asks of AI in implant design the question we ask of our own work: has a method been evaluated on independent data and in clinical use, or only on a benchmark? Here the answer is mostly the latter, and the review says so.

From Eden Health, Vlad Tau was responsible for the review's software, took part in the investigation and data curation and co-wrote the first draft; Alex Lucaci contributed to the methodology and resources. Every use of AI in the work is disclosed in the paper and logged in its supplementary materials.

What the study asked

In clinical practice, many patients at risk deny or minimise suicidal thoughts during assessment; the paper cites denial rates of roughly half among people with ideation. Interviews and rating scales depend on what a person is willing to say, so this dissimulation problem limits how well clinicians can judge current risk.

The study asked whether an objective signal, the voice, carries information about a current clinical state that does not depend on self-report, and whether a deep-learning model can read it.

How it was done

Speech recordings were collected from 88 adult psychiatric inpatients at two Romanian hospitals, the "Prof. Dr. Al. Obregia" Psychiatry Clinical Hospital in Bucharest and the "Elisabeta Doamna" Psychiatry Hospital in Galați: 59 patients admitted after a recent suicide attempt and 29 patients with severe depression without documented suicidal ideation.

Patient speech was isolated by automatic speaker diarisation with manual verification, cut into two-second segments and converted to 128-band mel-spectrograms. A multi-scale convolutional neural network with residual connections and spatial attention classified the segments, and each patient was classified by majority vote over their segments.

Evaluation used repeated patient-level stratified five-fold cross-validation with several random seeds, so no patient ever contributed recordings to both the training and the test side of a split.

What it found

• At patient level, the model reached a mean accuracy of 72.7% (± 6.4) and a ROC-AUC of 0.824 (± 0.054) across repeated cross-validation runs.
• Sensitivity was high, 88.3% (± 9.5), while specificity was modest and variable, 41.3% (± 25.0): the model rarely missed the attempt group but often flagged the depression group too.
• Sex-controlled analyses were reported honestly: a female-only analysis reached a ROC-AUC of 0.866 (± 0.123), whereas a repeated sex-matched analysis fell to 0.621 (± 0.156), so part of the signal may relate to sex imbalance between groups.
• The study received no external funding. It was approved by the ethics committees of both hospitals and all participants gave informed consent.

Why it matters

It is a proof of concept, not a clinical tool. It shows that a voice signal, measured objectively, correlates with a current clinical state that patients may not disclose. High sensitivity with modest specificity means it is best read as a prompt for clinician attention, not a verdict.

The pipeline behind this study is the foundation of our on-device voice analysis work, where inference runs on the phone and raw audio never leaves it.