inferenceCoreTemp.ts

Purpose

This module predicts core body temperature (in Celsius) from physiological and environmental inputs using one of several ONNX model variants, with a physics-based fallback for environments where the ONNX runtime is unavailable.

Its responsibilities include:

• Validating all required finite numeric inputs before inference
• Selecting the appropriate ONNX model (or baseline) based on the model argument
• Constructing a correctly ordered feature vector per model variant
• Loading and caching the ONNX session via getOnnxSession
• Running the session and extracting the scalar core temperature output
• Classifying the result into a heat-stress severity state
• Returning a structured result with the temperature, model identifier, and state

Invariants

Input Validation

All required inputs must be finite numbers before any model is invoked.

The following fields are validated unconditionally:

• wbgt
• gender
• bmi
• age
• daysActive14

Invalid inputs throw an Error with a descriptive field name.

Optional Input Enforcement

Each model variant enforces its own optional input requirements at runtime:

Missing required optional inputs return { coreTempC: 0, ... } with an error log rather than throwing.

Zero-Value Error Sentinel

coreTempC: 0 is used as the error sentinel value throughout.

Callers must treat a coreTempC of 0 as an inference failure, not a valid physiological reading.

Feature Vector Ordering

Each model variant uses a fixed, order-sensitive feature vector.

Incorrect ordering would silently produce wrong predictions; the ordering is maintained per model in code and must not be changed without retraining:

State Classification Thresholds

Variants

Model Selection

The model parameter selects the inference path:

• "baseline" — Physics-based formula; no ONNX session is created. Intended for use in Expo Go or test environments where the native runtime is unavailable.
• "watch-only" — ONNX model using heart rate and environmental features (filtered_model2.onnx).
• "hybrid" — ONNX model using both heart rate and accelerometer features (filtered_model3.onnx).
• "phone-only" (default) — ONNX model using accelerometer and environmental features (filtered_model1.onnx).

Baseline Formula

The fallback computation approximates core temperature as:

T = 36.9
  + 0.08 × (wbgt − 28)
  + 0.0002 × cpm40
  + 0.01 × ((bmi − 25) / 5)
  + 0.01 × ((age − 35) / 10)
  + genderAdj

Where genderAdj is 0.03 for male, 0.0 for female, and 0.01 for unknown.

The result is clamped to [35.0, 42.0] °C.

Exports

Gender

type Gender = 0 | 1 | 2;

Encodes biological sex used during model training:

CoreTempInputs

type CoreTempInputs = {
	wbgt: number;
	age: number;
	gender: Gender;
	bmi: number;
	daysActive14: number;
	cpm40?: number;
	hbpm40?: number;
};

All required fields must be finite numbers. Optional fields are enforced per model variant.

CoreTempResult

type CoreTempResult = {
	coreTempC: number;
	modelUsed: string;
	state: "normal" | "elevated" | "heat-strain" | "heat-stroke-risk";
};

coreTempC: 0 indicates an inference failure. modelUsed identifies which model path was taken.

runCoreTempInference(...)

async function runCoreTempInference(
	inputs: CoreTempInputs,
	model?: string,
): Promise<CoreTempResult>

Runs core temperature inference using the specified model variant.

Parameters

Returns

A Promise<CoreTempResult> containing the predicted core temperature, model identifier, and heat-stress state.

Throws

Throws when required inputs are non-finite (via validateInputs). Returns a zero-sentinel result (does not throw) when optional model-specific inputs are missing or when no model could be selected.