Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PI-NAIM: Path-Integrated Neural Adaptive Imputation Model

Afifa Khaled, Ebrahim Hamid Sumiea

TL;DR

PI-NAIM addresses missing data in multimodal settings by dynamically routing samples to either MICE or GAIN with temporal analysis, guided by missingness context and complexity. It fuses pathway outputs with cross-path attention and trains in an end-to-end, curriculum-guided framework that balances imputation fidelity and downstream task performance via uncertainty-aware fusion. Empirical results on MIMIC-III show RMSE $0.108$ and AUROC $0.812$ for mortality, with additional gains in CIFAR-imputation fidelity, demonstrating robust performance across temporal and visual domains. The modular architecture enables seamless integration into real-world pipelines handling incomplete sensor measurements and missing modalities, with publicly available code to support reproducibility and adoption.

Abstract

Medical imaging and multi-modal clinical settings often face the challange of missing modality in their diagnostic pipelines. Existing imputation methods either lack representational capacity or are computationally expensive. We propose PI-NAIM, a novel dual-path architecture that dynamically routes samples to optimized imputation approaches based on missingness complexity. Our framework integrates: (1) intelligent path routing that directs low missingness samples to efficient statistical imputation (MICE) and complex patterns to powerful neural networks (GAIN with temporal analysis); (2) cross-path attention fusion that leverages missingness-aware embeddings to intelligently combine both branches; and (3) end-to-end joint optimization of imputation accuracy and downstream task performance. Extensive experiments on MIMIC-III and multimodal benchmarks demonstrate state-of-the-art performance, achieving RMSE of 0.108 (vs. baselines' 0.119-0.152) and substantial gains in downstream tasks with an AUROC of 0.812 for mortality prediction. PI-NAIM's modular design enables seamless integration into vision pipelines handling incomplete sensor measurements, missing modalities, or corrupted inputs, providing a unified solution for real-world scenario. The code is publicly available at https://github.com/AfifaKhaled/PI-NAIM-Path-Integrated-Neural-Adaptive-Imputation-Model

PI-NAIM: Path-Integrated Neural Adaptive Imputation Model

TL;DR

PI-NAIM addresses missing data in multimodal settings by dynamically routing samples to either MICE or GAIN with temporal analysis, guided by missingness context and complexity. It fuses pathway outputs with cross-path attention and trains in an end-to-end, curriculum-guided framework that balances imputation fidelity and downstream task performance via uncertainty-aware fusion. Empirical results on MIMIC-III show RMSE and AUROC for mortality, with additional gains in CIFAR-imputation fidelity, demonstrating robust performance across temporal and visual domains. The modular architecture enables seamless integration into real-world pipelines handling incomplete sensor measurements and missing modalities, with publicly available code to support reproducibility and adoption.

Abstract

Medical imaging and multi-modal clinical settings often face the challange of missing modality in their diagnostic pipelines. Existing imputation methods either lack representational capacity or are computationally expensive. We propose PI-NAIM, a novel dual-path architecture that dynamically routes samples to optimized imputation approaches based on missingness complexity. Our framework integrates: (1) intelligent path routing that directs low missingness samples to efficient statistical imputation (MICE) and complex patterns to powerful neural networks (GAIN with temporal analysis); (2) cross-path attention fusion that leverages missingness-aware embeddings to intelligently combine both branches; and (3) end-to-end joint optimization of imputation accuracy and downstream task performance. Extensive experiments on MIMIC-III and multimodal benchmarks demonstrate state-of-the-art performance, achieving RMSE of 0.108 (vs. baselines' 0.119-0.152) and substantial gains in downstream tasks with an AUROC of 0.812 for mortality prediction. PI-NAIM's modular design enables seamless integration into vision pipelines handling incomplete sensor measurements, missing modalities, or corrupted inputs, providing a unified solution for real-world scenario. The code is publicly available at https://github.com/AfifaKhaled/PI-NAIM-Path-Integrated-Neural-Adaptive-Imputation-Model

Paper Structure

This paper contains 35 sections, 11 equations, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. PI-NAIM Architecture
  4. Input Processing
  5. Binary Mask $M$
  6. Missingness Embeddings $E$
  7. Dynamic Path Selection
  8. Missingness Rate (MR) as a Scalar Proxy
  9. MICE Path: High-Efficiency Statistical Branch
  10. GAIN Path with Temporal Analysis
  11. Cross-Path Attention Fusion
  12. Task-Supervised Adaptive Fusion
  13. Training Strategy
  14. Curriculum Masking
  15. Multi-Task Learning with Adaptive Weighting
  16. ...and 20 more sections

Figures (2)

  • Figure 1: PI-NAIM end-to-end architecture flowchart illustrating the four main stages: (1) input initialization and missingness embedding, (2) dynamic routing and imputation through MICE or GAIN paths based on missingness rate, (3) adaptive fusion combining imputed and task representations via cross-path attention, and (4) output generation including downstream task training, prediction, and uncertainty quantification. The design enables efficient, context-aware handling of diverse missingness patterns across temporal and multimodal data.
  • Figure 2: a and b architectural comparison and holistic summary of model capabilities, c training metrics on the MIMIC admissions data, showing the joint loss $\mathcal{L}_{\text{imp}} + \mathcal{L}_{\text{task}} + \mathcal{L}_{\text{reg}}$ and d illustration of the curriculum masking strategy's effectiveness.