AI tool cuts unexpected deaths in hospital by 26%, Canadian study finds

Original Article ↗ Hacker News Discussion ↗

Type of “AI” and What It Actually Does

  • Tool is based on a time‑aware Multivariate Adaptive Regression Splines (MARS) model, not an LLM.
  • Many see it as “classical” statistics / machine learning rather than novel AI.
  • Critics say headline uses “AI” as marketing; paper itself mainly uses “machine learning.”
  • Supporters argue model-based risk prediction integrating multiple labs over time is meaningfully beyond simple thresholds.

Effect Size: Relative vs Absolute Risk

  • Reported 26% reduction is relative risk (2.1% → 1.6% mortality).
  • Absolute risk reduction is ~0.69%, with an estimated number needed to treat (NNT) of ~156.
  • Some argue this small absolute gain plus 2:1 false positives makes clinical value modest.
  • Others counter that saving 1 life per ~156 patients is meaningful, especially if costs are low.

Alerts, False Positives, and Alarm Fatigue

  • Model accepts ~2 false alarms per true alarm; some find this reasonable prioritization in understaffed wards.
  • Others worry high false‑positive rates will drive “alert fatigue” and ignored warnings, or trigger unnecessary tests/interventions with their own risks.
  • Success depends heavily on workflow integration and how easy it is for staff to see and act on alerts.

Staffing, Incentives, and System Design

  • Many see the tool as compensating for nurse/doctor understaffing and delayed lab review.
  • Debate over whether such efficiency gains will improve care or justify further resource cuts (“just good enough” equilibrium).
  • Discussion contrasts Canadian single‑payer/non‑profit hospitals with US for‑profit systems, but notes cost‑cutting and bureaucracy exist in both.

Definitions of AI and Hype

  • Long debate on what counts as “AI”: simple rules vs regression vs ML vs LLMs.
  • Some want to reserve “AI” for modern neural/LLM systems; others see any approximate reasoning under uncertainty as AI.
  • Several commenters stress that simple, transparent ML often outperforms complex “shiny” models in healthcare.

Patient Experience and Advocacy

  • Multiple comments emphasize that hospital care quality still hinges on human factors: understaffed, burned‑out nurses and doctors.
  • Strong theme that having a family advocate at the bedside remains crucial, regardless of predictive tools.