Diffmv: A Unified Diffusion Framework for Healthcare Predictions with Random Missing Views and View Laziness

TL;DR

Diffmv tackles two core issues in multi-view healthcare prediction: random missing EHR views and view laziness. It unifies multi-view imputation within a diffusion-denoise pipeline using a binary mask and two contextual conditions (intra- and inter-patient) to sharpen missing-view generation, while a classifier-free, inverse-advantage reweighting balances view usage during optimization. The approach yields superior predictions across phenotype and LOS tasks on MIMIC-III, eICU, and MIMIC-IV-Note, outperforming state-of-the-art baselines and showing robust performance under varying missingness. By imputing complete EHRs and leveraging longitudinal representations, Diffmv enhances predictive reliability and offers interpretable, distribution-aligned generations with potential for broader clinical deployment.

Abstract

Advanced healthcare predictions offer significant improvements in patient outcomes by leveraging predictive analytics. Existing works primarily utilize various views of Electronic Health Record (EHR) data, such as diagnoses, lab tests, or clinical notes, for model training. These methods typically assume the availability of complete EHR views and that the designed model could fully leverage the potential of each view. However, in practice, random missing views and view laziness present two significant challenges that hinder further improvements in multi-view utilization. To address these challenges, we introduce Diffmv, an innovative diffusion-based generative framework designed to advance the exploitation of multiple views of EHR data. Specifically, to address random missing views, we integrate various views of EHR data into a unified diffusion-denoising framework, enriched with diverse contextual conditions to facilitate progressive alignment and view transformation. To mitigate view laziness, we propose a novel reweighting strategy that assesses the relative advantages of each view, promoting a balanced utilization of various data views within the model. Our proposed strategy achieves superior performance across multiple health prediction tasks derived from three popular datasets, including multi-view and multi-modality scenarios.

Figures (12)

• Figure 1: (a)-(b) refer to random missing view and missing status across three datasets. (c) Phenotype (PHE) prediction for patients with at most one view missing (One-miss) and two views missing (Two-miss).
• Figure 2: Overview of Diffmv. We first vectorize the EHR data for each view (V) and apply forward diffusion to generate noisy vectors. These vectors are then concatenated and fed into a unified denoising network for reverse denoising, using a binary mask, historical data, and prototypes as conditions. Next, we utilize the trained diffusion framework to impute missing views and pass the EHRs to a sequential model (SEQ), followed by inverse advantage (Adv) reweighting to refine predictions. "Enc" refers to visit encoding in Eq.\ref{['eq:1']}-\ref{['eq:2']}, while "Dec" denotes the rounding process in Eq. \ref{['eq:9']}. Both $q$ and $p$ signify conditional probability distributions, distinguishing between the forward and reverse processes. $\mathcal{L}_{\text{vlb}}$ and $\mathcal{L}_{\text{inv}}$ refer to Eq. \ref{['eq:14']} and Eq. \ref{['eq:18']}.
• Figure 3: Warm-cold examination. To enhance the figure's appearance, MedDiffusion is abbreviated as MedDiff.
• Figure 4: Group Analysis. G1-G4 represent patients with $(\frac{1}{2}, \frac{2}{3}]$, $(\frac{1}{3}, \frac{1}{2}]$, $(\frac{1}{6}, \frac{1}{3}]$,$[0, \frac{1}{6}]$ missing view ratio. Please note that $\frac{2}{3}$ represents the maximum limit for missing ratio, as at least one view is retained for each visit.
• Figure 5: Margin distribution generation. For Figure \ref{['fig:case:jsd']}, we present the frequency distributions of generated and real entities from the medication view. For Figure \ref{['fig:case:heat']}, We randomly select 50 missing views.