{
  "field_id": "F032",
  "field": "clinical machine learning",
  "directions": [
    {
      "direction_id": "G0",
      "name": "Assessment of current clinical and mental-health states",
      "scope": "Determine current disease, lesions, symptoms, or severity from existing patient evidence, including medical images, physiological and behavioral signals, structured clinical features, and mental-health cues in natural self-expression."
    },
    {
      "direction_id": "G1",
      "name": "Structuring documented clinical information",
      "scope": "Turn clinical narratives or records into usable structured clinical information, including extracting documented facts and assigning standardized diagnostic or procedural codes."
    },
    {
      "direction_id": "G2",
      "name": "Forecasting patient outcomes and disease courses",
      "scope": "Infer patients’ subsequent states from current or past evidence, including event risk and timing, neurodegenerative disease course, trajectories of repeated measurements, and future outcomes derived from images."
    },
    {
      "direction_id": "G3",
      "name": "Clinical learning with deficient predictive evidence",
      "scope": "Maintain effective inference for target patients and rare outcomes when evidence available for clinical prediction is incomplete or insufficiently represented, including the distinct gaps of missing inputs at prediction time and scarce outcome classes during training."
    },
    {
      "direction_id": "G4",
      "name": "Learning from heterogeneous clinical evidence",
      "scope": "Learn to align, fuse, or represent heterogeneous clinical information so it supports patient-level inference or can be reused across multiple clinical tasks."
    },
    {
      "direction_id": "G5",
      "name": "Validity of clinical models across uses and populations",
      "scope": "Assess or improve the trustworthiness of clinical-model conclusions for their intended uses, including changes in deployment conditions, disparities across patient groups, and evaluation of targets, metrics, stability, and workflow performance."
    },
    {
      "direction_id": "G6",
      "name": "Inspectable patient-level predictions",
      "scope": "Make individual patient predictions inspectable and suitable for appropriate use by assessing their confidence and need for review or explaining the clinical evidence supporting them."
    },
    {
      "direction_id": "G7",
      "name": "Personalized intervention decisions",
      "scope": "Compare the treatment benefits patients might receive or learn policies for choosing single or sequential interventions based on patient state."
    },
    {
      "direction_id": "G8",
      "name": "Discovery of clinical phenotype structure",
      "scope": "Discover and clinically characterize patient subgroups, symptom combinations, or latent disease patterns without making individual supervised prediction the principal output."
    },
    {
      "direction_id": "G9",
      "name": "Clinical learning without pooling patient data",
      "scope": "Train clinical prediction models across data holders when patient-level data cannot be pooled centrally."
    }
  ]
}