HALO-X: Multi-Modal RPM in Cancer Care

• HALO-X Platform is a multi-modal remote patient monitoring system that integrates wearable sensor data, patient-reported outcomes, clinical event codes, and demographics into unified chronological records.
• It employs a modular encoding pipeline including continuous variable encoders, event embeddings, and explicit missingness representations processed by Transformer layers to handle asynchronous data.
• The platform demonstrates clinical utility with an AUROC of 0.70 in risk prediction, enabling proactive interventions in cancer care.

The HALO-X Platform is a multi-modal remote patient monitoring (RPM) data infrastructure and AI modeling ecosystem designed for application in cancer care. It integrates heterogeneous continuous and event-based patient observations—including wearable sensor streams, patient-reported outcomes, clinical events, and static demographic data—into unified chronological records for use by multi-modal, sequence-based AI models. The platform enables continuous risk forecasting of adverse clinical events, and supports robust handling of real-world asynchronous and missing data typical in longitudinal medical settings (Liu et al., 30 Nov 2025).

1. Data Modalities and Streaming Aggregation

HALO-X collects and processes data across several axes:

• Wearable Device Streams: Continuous physiological signals such as heart rate and step count, processed into daily summaries (e.g., daily maximum heart rate, total step count).
• Patient-Reported Surveys: Quality of Recovery (QoR-15) scores and wellness check-in binaries, sampled daily or less frequently.
• Clinical Events: Time-stamped codes for treatments (chemotherapy, immunotherapy), inpatient and outpatient visits, emergency department encounters, and delays in care.
• Demographics: Static features (age, gender, BMI) encoded via feed-forward networks.
• Missingness Indicators: Fraction of time wearable devices are worn per day, duration of observation gaps, and explicit missingness tokens for incomplete entries.

These sources are preprocessed into normalized vectors and projected into token representations of fixed dimension. Each day's record comprises a concatenation of observed tokens, facilitating chronological modeling.

2. Multi-modal Tokenization and Temporal Modeling

HALO-X implements a modular encoding pipeline:

• Continuous Variable Encoder (CVE): Processes wearable and survey features via linear projection of clipped and normalized values into D-dimensional token encodings.
• Clinical Event Embedding: Embeds one-hot event type vectors using dedicated learnable embedding tables.
• Missingness Token Generator: Creates tokens that represent percent wearable usage and absence duration, capturing non-random missingness (MNAR) patterns.
• Static Demographic Encoder: Feeds static covariates through a feed-forward layer to produce a baseline patient code.

All tokens are augmented with time-step embeddings and modality-type tags:

$$z_{d,m} = e^{(m)}_{d} \;+\; \mathrm{TypeEmbed}(m)\;+\; p_d$$

Here, $d$ is the day index, $m$ indexes the feature modality, and $e^{(m)}_{d}$ denotes the modality-specific token at day $d$.

The resulting token sequence is handled by a stack of Transformer layers performing multi-head self-attention, with cross-modal information exchange enabled implicitly through co-attention mechanisms (Liu et al., 30 Nov 2025):

$$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\bigl(\tfrac{QK^\top}{\sqrt{d_k}}\bigr)V$$

Tokens for unobserved modalities are omitted; no padding or forced imputation is applied beyond window truncation.

3. Asynchronous, Incomplete Data Handling

Recognizing the asynchronous nature of RPM data, HALO-X incorporates missingness representation and accommodates variable-length chronological records:

• Sparse Observation Tensor: Only observed data points are tokenized; time steps or modalities missing observations are dropped.
• Explicit Missingness Encoding: Includes daily wear percent and absence duration tokens to inform the model about observation reliability, without imputation bias.
• Transformer Windowing: Token sequences are windowed up to 1000 sequential steps, handling protracted monitoring periods and enabling sliding-window forecasting.

This architecture directly models temporality, enabling response to gaps, device drop-off, and survey non-compliance.

4. Risk Forecasting and Labeling Approach

HALO-X's predictive target is a continuous forecast of future adverse clinical events:

• Binary Labeling: For each sliding window, the target label is set as 1 if any adverse clinical event occurs within the next 28 days, otherwise 0.
• Classification Head: A [CLS] token serves as a pooled representation after Transformer layers; its output is projected by a linear layer and mapped via a logistic sigmoid to yield risk probabilities:

$$\hat{y} = \sigma(w^\top h_{\text{cls}} + b), \quad \sigma(x)=\frac{1}{1+e^{-x}}$$

• Loss Function: Training is supervised via binary cross-entropy with L2 regularization:

$$\mathcal L = -\frac{1}{B}\sum_{i=1}^B\left[y_i\log\hat y_i + (1-y_i)\log(1-\hat y_i)\right] + \lambda\|\theta\|_2^2$$

• Sampling: Training data is generated via sliding windows: each window uses at least 14 historical days to predict the subsequent 28-day risk.

5. Model Training, Optimization, and Evaluation

• Training Setup: Batches of 128 are run for 80 epochs with Adam optimizer (lr=5e–4), stratified by patient to prevent leakage.
• Class Imbalance: No explicit weighting; AUROC is the preferred metric due to negative:positive ratio ≈6.4:1.
• Performance Metrics: Empirically, HALO-X achieves an accuracy of 83.9% (95% CI: [81.5%, 86.3%]) and AUROC of 0.70 (95% CI: [0.64, 0.76]) for risk prediction over 2.1M data points and 6,080 patient-days in 84 cancer patients.

6. Attention-Based Feature Importance Extraction

Feature importance is calculated by aggregating mean attention weights from the final Transformer layer (attention between [CLS] and feature tokens). Top contributors (with approximate importance values) include prior chemotherapy (0.085), A&E visits (0.072), prior immunotherapy (0.065), wellness check-ins (0.058), sleep and work survey items, daily max heart rate (0.038), and percent wearable usage (0.033). This reveals which multi-modal features are most informative for anticipating adverse events.

7. Case Study and Clinical Utility

A timeline analysis illustrates HALO-X's operational mode: periods of stable low risk (ŷ≈0.10–0.20) are followed by rising risk scores in response to missed wear days and clinical events, culminating in risk peaks that precede actual adverse outcomes by several days. This early warning capacity demonstrates a plausible implication for proactive patient support and timely intervention, as elevated risk is flagged prior to event occurrence.

8. Technological and Clinical Significance

The HALO-X platform demonstrates pragmatic multi-modal data fusion, fine-grained temporal modeling, and robust tolerance of asynchronous/missing data. It establishes feasibility for AI-powered RPM in complex cancer care environments, providing automated, longitudinal risk forecasting. The platform's modeling architecture and empirical results set a precedent for further research on clinical-grade remote monitoring using heterogeneous, real-world signal streams (Liu et al., 30 Nov 2025).