SubHealthAI: Predictive and Explainable AI for Early Detection of Subclinical Health Decline

1. Background and Motivation

Chronic disease accounts for about 90% of U.S. healthcare spending according to the CDC. Most conditions such as diabetes, cardiovascular disease, and metabolic dysfunction develop silently for years before diagnosis. Explainable machine learning methods such as SHAP can improve transparency and trust for preventive decision-making [2]. Current consumer health tools track steps or sleep but do not combine multiple signals or highlight subtle trends that suggest early dysfunction. Clinicians often receive fragmented data that cannot be interpreted easily or integrated into preventive care workflows. Prior studies have explored explainable anomaly detection and wearable-based health prediction [3,4], but most target post-diagnosis monitoring or lack interpretable rationale for pre-diagnostic warning. SubHealthAI addresses this gap by integrating predictive modeling with human-interpretable explainability.

SubHealthAI was created to fill that gap. It converts raw wearable and lifestyle data into structured early-warning signals that are explainable and auditable. The goal is not diagnosis, but timely awareness and preventive guidance.

2. System Overview

The system brings together data from wearables, lifestyle logs, and future lab reports.

• Inputs: heart rate, heart rate variability, resting heart rate, sleep duration, and daily activity.
• Processing: each metric is compared with the user’s personal baseline to detect anomalies or sustained changes.
• Outputs: a color-coded risk indicator, a short plain-language summary, and a clinician PDF that lists the main contributing factors.

SubHealthAI’s dashboard shows a weekly trend line, daily risk score, and a simple explanation of which patterns triggered the change. Each report includes a disclaimer stating that it is not diagnostic and should be interpreted under clinical oversight.

As shown in Figure 1, the system integrates ingestion, anomaly detection, forecasting, explainability, and audit logging.

Figure 2 details the data flow and report generation pipeline.

Figure 3 shows the dashboard view used to surface risk and explainability to users.

3. Technology Summary

SubHealthAI uses two lightweight machine learning models. The first detects anomalies in daily health metrics using an isolation forest. The second predicts the next day’s risk level using a small sequence model. Results are stored in a secure database with complete audit logs.

Each prediction is paired with an explainability chart that shows how features like sleep, heart rate, and activity influenced the score. Visuals are generated with SHAP and displayed on the dashboard and clinician report. All outputs and model versions are logged in the audit table for reproducibility and compliance. SHAP values provide per-feature contribution scores that explain each prediction in human-interpretable terms [2].

Figure 4 illustrates feature-level contributions to a sample prediction.

Future versions will include optional lab data such as C-reactive protein (CRP), hemoglobin A1c (HbA1c), and estimated glomerular filtration rate (eGFR) to detect inflammation or metabolic imbalance earlier. These values, combined with wearable trends, will help identify subtle health drift before a disease state appears.

4. Results

A test run using 14 days of simulated wearable data demonstrated stable risk calculations and consistent feature explanations. The system correctly flagged patterns such as low sleep, high resting heart rate, and reduced HRV, producing a clear risk summary with transparent reasoning. Explainability plots and clinician reports were stored securely for review and audit.

Figure 5 presents an example clinician-facing risk report generated by the pipeline.

4.1. Data and Preprocessing

We evaluated a 14-day prototype using synthetic wearable data modeled after patterns from open datasets (e.g., WESAD, MIMIC-IV). The dataset contained four core metrics—heart rate, HRV, sleep duration, and daily steps—augmented with Gaussian noise and time-warping to emulate realistic day-to-day variability. Features were normalized to z-scores; outliers were capped using interquartile range filtering; and rolling daily averages were computed to reduce noise.

4.2. Evaluation and Results

The Isolation Forest anomaly detector achieved precision $=0.81$ and recall $=0.76$ on simulated deviations from personal baselines. The GRU-based model achieved RMSE $=0.12$ and MAE $=0.09$ for next-day risk prediction on normalized scores. These preliminary results suggest stable behavior for surfacing early deviations with interpretable rationales while keeping the model lightweight and auditable.

5. Discussion

Most existing health tools focus on data tracking, not pattern interpretation. Apple Health, Fitbit, and WHOOP provide graphs but do not explain why changes occur or link them to actionable insights. Academic projects such as DeepPatient and BioWear predict health outcomes from electronic records but are not transparent or deployable for individual users.

SubHealthAI combines three elements that are rarely present together:

• Individualized baselines rather than population averages.
• Transparent explanations for each prediction.
• Compliance-ready audit logs and disclaimers.

This design makes it both practical for public use and safe for clinician oversight.

Most commercial health systems either prioritize consumer engagement or academic accuracy, but not explainable real-time inference. SubHealthAI bridges this divide by combining anomaly detection, short-horizon forecasting, and transparent SHAP visualization into a single preventive-care workflow.

While these results are based on synthetic data, they validate the core design principles of the system: interpretable anomaly detection, short-horizon forecasting, and auditable reporting. Future validation will use larger real-world datasets (e.g., MIMIC-IV, UK Biobank) and compare against baselines such as LSTM, ARIMA, and Prophet to quantify incremental value over standard time-series approaches.

6. Public Health Impact

If scaled nationally, SubHealthAI could support earlier interventions and reduce the long-term cost of chronic disease. Even a small percentage reduction in delayed diagnoses would save billions in healthcare spending each year.

The project aligns with current priorities of the U.S. Department of Health and Human Services (HHS), the Centers for Disease Control and Prevention (CDC), and the National Institutes of Health (NIH), which emphasize preventive care and digital health innovation.

By providing explainable, privacy-respecting early warnings, SubHealthAI directly supports these public health goals. The explainability-first design also aligns with ongoing federal emphasis on transparency and accountability in AI-enabled health tools.

SubHealthAI is designed as a low-cost, cloud-deployable model that could integrate into digital health programs under HHS preventive-care grants or NIH pilot frameworks.

7. Future Work

Planned features include:

• A refined architecture (Figure 1) including a Preventive Copilot LLM and multimodal lab/imaging extensions.
• Integration of lab data such as urinalysis, lipid panels, and vitamin D levels.
• Support for allergy reports and imaging summaries (e.g., MRI findings).
• A conversational assistant that helps users interpret results in natural language.
• Federated learning for private, personalized model updates.

We will integrate a lightweight agentic LLM copilot using minimal-chat orchestration (streaming prefill/decode with KV-cache), JSON tool contracts for safe calls to scoring/SHAP routines, and sandbox guardrails. The explainability UI will evolve from static plots to dynamic, WHOOP/Apple-style health graphs. Additional modalities (medication, jet lag, circadian disruption) will be incorporated to improve temporal context and personalization.

Figure 6 outlines the planned multimodal extension.