Language of Order: A Neural Network Translator for Medical Data ✛ ✎ ✦

Original: "Deep Multitask Learning for Mixed-Type Outcomes with Shared Sparsity"

· Huichao Li, Tong Wang, Sanguo Zhang, Shuangge Ma

arXiv:2607.00995 · 2026-07-01 · arXiv non-exclusive · independent analysis · 1 min

A new approach teaches a neural network to translate the language of numbers, diagnoses, and categories into the universal language of order — without manual tuning.

A medical record is a mix of numbers, diagnoses, and categories. A typical neural network (a program that learns) gets lost because it can't match protein levels (numbers) with diagnoses (yes/no) and tumor subtypes (categories). Their errors are incomparable, like a translator confusing centimeters with kilograms. The authors abandoned exact values and translated everything into the language of order. The program simply asks: is this patient's indicator higher or lower than the neighbor's? Such ranking makes errors commensurate.

First, a common block of the network sifts through hundreds of genes, leaving only those important for all tasks at once. A special technique (it zeros out the contribution of useless genes) removes the noise. Then, for each task, the network outputs conditional scores and compares them by order. The questions 'Is the diagnosis correct?' and 'Is the protein level accurate?' reduce to one: was the order among patients guessed correctly?

On 234 lung cancer patients, accuracy rose from 0.951 to 0.968—per thousand patients, that's 17 saved from false diagnoses. Among the selected genes, not only lung-related ones: E2F1 (previously linked to breast cancer) and Scd1 (fat metabolism). Thus, the 'order translator' finds common ground between immunity and the tumor.

The idea of ranks is almost 40 years old: a criterion from 1987, a smooth curve from 2007. For the first time, they are combined in a single neural network.

Whole graph →

The rigid 'greater/less' rule was replaced with a smooth S-shaped curve so the computer could learn gradually, without jumps. And the idea of comparing orders was proposed back in 1987.

It's like a universal translator from sci-fi movies, only instead of languages, it's types of medical data.

Y_{t} = D_{t}(f_{t}(X) + \epsilon_{t})

The measured response Y_t is obtained from a latent smooth function f_t(X) with noise, passed through an unknown monotonic transformation D_t (e.g., a logarithm that does not distort order).

s(\zeta; a_n) = \frac{1}{1+\exp(-\zeta/a_n)}

The smooth sigmoid replaces the hard 'greater-or-less' indicator, where the parameter a_n controls the steepness of the transition.

Scientists

Andrey Markov

Concepts

transfer learning training training data predictive learning neural networks genetic algorithm random forest Large Language Models decision making boosted decision tree generalization xgboost probabilistic model regularization methods prediction error gibbs sampling metropolis hastings algorithm overfitting weighted mean bayesian hierarchical model jackknife resampling

I am the author

Language of Order: A Neural Network Translator for Medical Data ✛ ✎ ✦