Table 2 Inputs and outputs for our per-signal upstream LSTMs.

From: Forecasting adverse surgical events using self-supervised transfer learning for physiological signals

E

Domain

Range (Upstream Task)

\({{{{\mathcal{L}}}}}_{E}\)

rand

\({X}_{{d}_{j}}^{i}[t-59:t]\in {{\mathbb{R}}}^{60}\)

\({{\emptyset}}\)

\({{\emptyset}}\)

auto

\({X}_{{d}_{j}}^{i}[t-59:t]\in {{\mathbb{R}}}^{60}\)

\({X}_{{d}_{j}}^{i}[t-59:t]\in {{\mathbb{R}}}^{60}\)

MSE

next

\({X}_{{d}_{j}}^{i}[t-59:t]\in {{\mathbb{R}}}^{60}\)

\({X}_{{d}_{j}}^{i}[t+1:t+5]\in {{\mathbb{R}}}^{5}\)

MSE

min

\({X}_{{d}_{j}}^{i}[t-59:t]\in {{\mathbb{R}}}^{60}\)

\(\min(\mathop{X}\nolimits_{{d}_{j}}^{{i}}[t+1:t+5])\in {{\mathbb{R}}}^{1}\)

MSE

hypo

\({X}_{{d}_{j}}^{i}[t-59:t]\in {{\mathbb{R}}}^{60}\)

yi ∈ {0, 1}

BCE

  1. We denote embedding names in italics.