Using 233 de-identified 12-lead ECGs from the PTB Diagnostic ECG Database and MIMIC-IV-ECG (198 with bipolar disorder, major depressive disorder, or schizophrenia; 35 healthy controls), an ensemble ML model classified psychiatric disorders vs controls with 94.8% accuracy (95% CI: 92.1–96.8%), including 89.7% sensitivity for major depressive disorder (MDD). The study also reduced 1,248 ECG-derived features to 84 optimal features and reported ~12.7 seconds processing time per participant on standard workstations.

Why It Matters To Your Practice

• ECG-based ML suggests a potential adjunctive, objective signal for psychiatric classification using data already common in many health systems.

• Performance was reported as stable across age groups (93.8%–95.2%) and resilient to noise (maintained >85% accuracy down to 15 dB SNR), which matters for real-world signal quality.

• Fast runtime (~12.7 ± 2.3 s per patient) lowers operational friction if integrated into clinical workflows.

Clinical Implications

• This is not a standalone diagnostic test: it would need to be positioned as a decision-support tool alongside clinical interview, collateral, and standardized rating scales.

• Condition-specific sensitivity varied (bipolar 92.3%, MDD 89.7%, schizophrenia 95.1%), so clinicians should expect uneven performance by diagnosis and potential spectrum effects.

• Medication and comorbidity confounding (e.g., antipsychotics, antidepressants, substance use, cardiac disease) remains a key unanswered question before clinical adoption.

Insights

• Feature reduction from 1,248 to 84 via recursive feature elimination suggests many ECG markers may be redundant, and a smaller feature set could support interpretability and deployment.

• Validation relied on stratified 10-fold cross-validation with bootstrap CIs (1,000 iterations) within a relatively small dataset (n=233), raising the usual generalizability questions for multi-class psychiatric ML.

• Using public repositories (PTB + MIMIC) is a strength for reproducibility, but prospective, multi-site testing is still required to assess drift and real-world prevalence.

The Bottom Line

• A multi-lead ECG ML system trained on PTB and MIMIC-IV-ECG data reported 94.8% accuracy for distinguishing major psychiatric disorders from healthy controls, with MDD sensitivity of 89.7%.

• For clinicians, the near-term takeaway is “promising adjunct signal” — not “replacement for diagnosis” — pending prospective validation and confounding control.