The pipeline at a glance
ML4HF™ runs a two-stage pipeline on the invasive hemodynamic values a clinician already has from a Swan-Ganz, PiCCO, CardioMEMS, or right-heart cath.
Stage 1 reconstructs a patient-specific pressure-volume (PV) loop and computes a panel of advanced energetic indices that quantify contractile reserve, ventricular-arterial coupling, and circulatory work.
Stage 2 is a machine-learning predictor trained on annotated advanced heart-failure outcomes. It returns a one-year composite event-rate estimate (death · LVAD · transplant) with an explicit uncertainty band.
The receiving clinician sees both the synthesis (Stevenson, SCAI, ACC/AHA classifications + energetic indices) and the underlying physiology (the PV loop and raw data). The reasoning is transparent — never a black box.
Stage 1 — Pressure-volume reconstruction
Pressure-volume mechanics are the gold standard for characterising cardiac function but are rarely computed at the bedside because the math is non-trivial. ML4HF™ fits a model-based PV loop to the patient's invasive inputs using established cardiovascular mechanics, then derives the indices clinicians need.
Indices computed
- LV MPS — Left Ventricular Myocardial Performance Score; index of overall contractile reserve.
- LV CPO — Left Ventricular Cardiac Power Output (Watts); strong correlate of mortality in shock.
- LV VAC — Ventricular-Arterial Coupling Ratio (Ees/Ea); efficiency of energy transfer.
- API — Aortic Pulsatility Index; integrates systemic afterload responsiveness.
- PAPi — Pulmonary Arterial Pulsatility Index; sentinel for right-ventricular dysfunction and RV-PA uncoupling.
- LV Efficiency — ratio of stroke work to total PVA, reflecting metabolic cost of each contraction.
- EesLV, EaLV, SWLV, PELV, PVALV, MVO₂LV — the underlying components of the PV-loop fit, available for clinicians who want to inspect the physiology directly.
These indices are computed instantly and surfaced alongside the staging classifications so clinicians can interpret them in context, not in isolation.
Stage 2 — Machine-learning predictor
The Stage 1 indices, together with the raw hemodynamic inputs, feed an ML predictor that estimates the patient's one-year risk of a composite endpoint: death, LVAD implantation, or heart transplant.
Endpoint
One-year composite (death · LVAD · transplant), aligned with the SHFM-style composite used widely in advanced HF outcome studies. This endpoint was selected because it captures the decisions advanced HF programs actually make — transplant evaluation, durable mechanical support consideration, palliative care discussion — rather than mortality alone.
Uncertainty quantification
The predictor returns a mean estimate and a standard-deviation band. Both are displayed prominently in the clinical snapshot. A wider band signals that the patient's physiology is in a less-represented region of training data and that the point estimate should be interpreted with extra caution.
Continuous refinement
The model improves continuously from de-identified cases clinicians contribute through the app's opt-in data sharing. No patient identifiers are ever collected; only the hemodynamic values and computed indices flow back to refine the cohort.
Classifications returned
Stevenson profile
The classic warm/cold × wet/dry quadrant: A (warm-dry, compensated), B (warm-wet, congested), L (cold-dry, hypoperfusion), C (cold-wet, classic cardiogenic shock physiology). Computed from cardiac index, PCWP, and clinical context.
SCAI shock stage
A (at risk) through E (extremis). Maps the patient's perfusion, biomarker, and circulatory support trajectory onto the 2019 Society for Cardiovascular Angiography and Interventions consensus framework.
ACC/AHA heart-failure stage
Stage A (at risk) through stage D (advanced HF), anchored to the 2022 AHA/ACC/HFSA guidelines. Drives the referral and management pathway surfaced in the app.
Key insights engine
Beyond raw classifications, ML4HF™ surfaces up to three narrative key insights per case — short clinical phrases that translate the indices into language a clinician can act on directly. Examples: "Severe hypoperfusion", "Reduced forward flow", "Elevated afterload", "RV dysfunction", "Energetic status critical", "Improving trajectory".
Each insight is derived from a transparent threshold rule over the computed indices (e.g., PAPi < 1.0 → RV-PA uncoupling) and is sorted by clinical severity. The full rule set is open in the source for any clinician who wants to audit how a particular insight was triggered.
Triage and escalation logic
The Rapid Triage tab and the page-2 referral PDF surface phenotype-driven management pathways and escalation triggers. Pathways are bucketed by time horizon (Initial / Reassess 6–12 h / 24–72 h) and use the deliberately non-directive language of decision support — "consider", "engage", "pathways consistent with this phenotype" — rather than prescriptive recommendations.
Escalation triggers use generic guideline thresholds (MAP < 60, CI < 1.8 sustained, PAPi < 0.7, UOP < 0.5 mL/kg/hr) rather than patient-individualised values, so the clinician retains responsibility for the threshold decision in context.
Training dataset
- Trained on a large cohort of patients with advanced and end-stage heart failure (SCAI Stages C–E predominant).
- Retrospective, IRB-approved hemodynamic + outcome data.
- Methodologically aligned with the energetic-tipping-point framework validated in Grinstein 2024 against a 30-day SCAI Stage C cohort.
- Continuously refined from de-identified contributions through the in-app opt-in.
- No patient identifiers (names, MRNs, room numbers, institutions) are collected at any point.
Limitations
- Predictions are cohort-trained probabilistic estimates. Individual patient trajectories may diverge.
- The model was trained predominantly on advanced HF / cardiogenic shock physiology. Generalisability to other populations (e.g., HFpEF without invasive monitoring, pediatric HF) has not been established.
- The training cohort is biased toward U.S. tertiary referral centers. Patients managed entirely in community settings may be under-represented.
- The PV-loop reconstruction is model-based and assumes physiologic plausibility. Extreme or implausible input combinations may produce non-physiologic loops; the app surfaces these as warnings.
- The 1-year composite endpoint captures hard outcomes but not quality of life, functional class, or symptom burden.
Regulatory positioning
ML4HF™ is not FDA-cleared. It is positioned as non-device Clinical Decision Support (CDS) under Section 520(o)(1)(E) of the U.S. Federal Food, Drug, and Cosmetic Act.
The receiving clinician is expected to independently review the underlying hemodynamic data and the cited basis for any recommendation, and to apply their own clinical judgment. ML4HF™ is not intended for emergency or time-critical decision-making.
The app surfaces the raw inputs, the computed indices, the staging classifications, and the cited literature — so the basis for any recommendation is auditable by the clinician at the point of use, not opaque behind an algorithm. This transparency is a deliberate design choice consistent with the FDA's 2022 final guidance on Clinical Decision Support Software.
If you are evaluating ML4HF™ for institutional use, the regulatory position is defensible as non-device CDS at present. We do not market ML4HF™ for diagnostic or treatment decisions, and we do not claim FDA clearance. Future commercial expansion may pursue 510(k) De Novo clearance.
Citations
The synthesis, classifications, and management pathways surfaced by ML4HF™ are aligned with the following peer-reviewed sources. Each links to PubMed for independent review.
- Stevenson LW. Tailored therapy to hemodynamic goals for advanced heart failure. Eur J Heart Fail. 1999. PubMed ↗
- Baran DA, Grines CL, Bailey S, et al. SCAI clinical expert consensus statement on the classification of cardiogenic shock. Catheter Cardiovasc Interv. 2019;94(1):29–37. PubMed ↗
- Heidenreich PA, Bozkurt B, Aguilar D, et al. 2022 AHA/ACC/HFSA Guideline for the Management of Heart Failure. J Am Coll Cardiol. 2022;79(17):e263–e421. PubMed ↗
- Forrester JS, Diamond G, Chatterjee K, Swan HJ. Medical therapy of acute myocardial infarction by application of hemodynamic subsets. N Engl J Med. 1976. PubMed ↗
- Grinstein J. Energetic tipping-point framework in advanced heart failure. Front Cardiovasc Med. 2024. PubMed ↗
Disclosure
ML4HF™ is developed by PACDynamic. The development team includes practising clinicians in advanced heart failure cardiology. No external funding has influenced the design, training, or marketing of the tool to date. Conflicts of interest, when any exist, will be disclosed alongside any peer-reviewed publication or institutional deployment of the platform.
For methodology questions, partnership inquiries, or institutional evaluation requests, contact rohan@pacdynamic.com.