402 ‒ NMR blood analysis: how mortality risk and more can be assessed from a single blood sample

KEYPOINTS

• 01

  NMR lipoprotein testing was born from debunking a flawed 1986 cancer test.

  A New England Journal paper claimed the width of a plasma NMR signal could diagnose any cancer. Otvos ran six leftover plasma samples; half looked 'cancerous,' but three came from women who had just given birth. The signal turned out to be a superposition of VLDL, LDL and HDL lipid signals. 'Cancer' samples simply had higher triglycerides and lower HDL, which narrowed the peak. He turned that composite signal into a way to count lipoprotein particles.

• 02

  Most reported LDL cholesterol is still estimated, not measured.

  Labs measure total and HDL cholesterol, then estimate VLDL cholesterol as triglycerides ÷ 5 (Friedewald) and subtract. Unless a report says 'direct,' LDL-C was probably calculated this way. Otvos helped develop a more accurate NIH equation used by LabCorp, but many labs still divide by five. Triglyceride assays measure glycerol, so rare free-glycerol states can falsely suggest high triglycerides.

• 03

  NMR counts particles because particle size shifts the signal frequency.

  Identical lipid methyl groups resonate at slightly lower frequency in larger particles. A 30-second spectrum is deconvoluted into contributions from each VLDL, LDL and HDL size subclass. Because NMR can't tell cholesterol from triglyceride inside, results are particle concentrations in nmol/L, not mg/dL.

• 04

  Small dense LDL's extra risk is explained by particle number, not particle size.

  At a given LDL-C, small LDL looks about threefold riskier. But smaller particles carry less cholesterol each, so the same LDL-C means more particles. Stratify by LDL-P instead, and size adds nothing. This matters because drugs that enlarge particles—niacin, fibrates, CETP inhibitors—failed outcome trials despite 'better' size profiles.

• 05

  'Large, fluffy' LDL is not benign.

  Asked about patients with LDL-P above 2,000 nmol/L who feel reassured by pattern A, Otvos calls it a specious idea. Familial hypercholesterolaemia patients have very high counts of large LDL particles; untreated homozygotes can die at 30–35.

• 06

  When LDL-C and LDL-P disagree, risk follows particle number.

  In Framingham (2007) and MESA (2011), Otvos ranked LDL-C and LDL-P as percentiles and called people within 12 percentile points 'concordant,' splitting the cohort roughly into thirds. Event curves were identical when the measures agreed. When LDL-P exceeded LDL-C, events rose faster; when cholesterol was high but particles low, events were lower. Discordance rises with each metabolic-syndrome criterion and is almost guaranteed with all five.

• 07

  LDL-P/apoB is a management tool more than a risk-prediction tool.

  Adding LDL-P to standard risk equations barely improves prediction, Otvos argues, because HDL-C in those models already captures the same signal indirectly. Where particles win is treatment: on statins, LDL-C often falls further than particle count. As LDL falls, triglyceride replaces cholesterol ester in the particle core, which can reveal residual risk and a case for adding drugs like PCSK9 inhibitors.

• 08

  LP-IR detects insulin resistance years before glucose rises.

  Insulin resistance is to diabetes what LDL is to atherosclerosis: the cause, integrated over time. Yet medicine waits for glucose, by which point roughly half of beta-cell function may be gone. LP-IR combines six lipoprotein size/subclass measures into a 0–100 score. In unpublished Diabetes Prevention Program data, it beat fasting insulin. Its fall with lifestyle change, and less with metformin, predicted lower diabetes incidence. The Diabetes Risk Index adds branched-chain amino acids.

• 09

  GlycA is a stable, cheap marker of systemic inflammation.

  The signal was first quantified just to subtract it from a fatty-acid peak. It comes from glycan decorations on acute-phase proteins, so it reflects both their levels and their glycosylation. Unlike hs-CRP, which can jump a thousandfold with infection and needs repeated measures, GlycA might double. It tracks IL-6 and falls with treatment of rheumatoid arthritis or psoriasis, but misses local inflammation.

• 10

  MVX predicts dying, more than getting a disease—'metabolic frailty.'

  The Metabolic Vulnerability Index combines GlycA and small HDL-P (the inflammation subscore, IVX) with citrate and the three BCAAs (the malnutrition subscore, MMX). Low BCAAs signal wasting, the opposite of the diabetes signal. Its predictive power holds across sex, obesity, diabetes, heart failure and kidney disease. Otvos argues composite fatal plus non-fatal endpoints make it look disease-specific, when it really reflects susceptibility to dying of whatever comes.

• 11

  The technology is at risk because diagnostics economics don't reward free information.

  One 150 µL plasma run yields a lipid panel, apoB via machine learning, glucose, LP-IR, GlycA and MVX at near-$1 marginal cost on $400–500k analyzers with no consumables. LabCorp bought LipoScience about a decade ago for lab revenue, not the in-vitro-diagnostics vision. It pulled the Vantera analyzers from places like Mayo and Cleveland Clinic, and the 15-year-old units are aging.