Modern Views of Machine Learning for Precision Psychiatry

In plain words

Psychiatry still diagnoses mostly from what patients report and what clinicians observe, and the same label can cover very different people. This review argues that machine learning (ML) and AI can help on several fronts: reading brain recordings, following people through phones and wearables, analysing speech, text and molecular measurements, and combining these sources into one picture of a person. It walks through the main ML approaches, the kinds of data used, and the open problems. Those problems include small samples, noisy labels, models that do not carry over to new settings, bias, privacy, and the gap between a pattern in a scan and a cause that a treatment could act on.

TL;DR

This is a narrative review, not a method paper. It surveys ML for neuroimaging, digital and mobile data, multimodal fusion, molecular phenotyping and explainable AI, and it ends with four case studies §1 §3.5. There is no model here to implement and no code. Use it to orient yourself and to find the cited primary work, and do not treat its numbers as reproduced results.

Problem

The authors argue that psychiatric diagnosis rests on symptom checklists and clinical judgement, which vary a lot between clinicians and patients §1 §3.1. Case-control designs compare patients with healthy people. They hide individual differences, and they struggle with the heavy overlap between diagnoses §3.1. The review frames the field around four issues: generalisability, interpretability, causality, and clinical integration §1. It says outright that its literature coverage is not exhaustive §1.

Method

The review has no method of its own. Its scope, by section:

• Neuroimaging basics. MRI, DTI, fMRI, PET, EEG, MEG, ECoG and fNIRS are described with their trade-offs. EEG and fMRI are named as the two most used modalities, and EEG is the cheaper and more portable one §2 fig. 1.
• Neuroimaging analyses. Task-related activation, structural and functional connectivity, brain dynamics, and multimodal neuroimaging §2.3.
• ML families (summarised in tab. 1):
  • Supervised learning, including logistic regression, SVMs, random forests, LASSO, connectome-based prediction, and relevance vector machines with automatic relevance determination. Multi-task and latent-space variants are covered §3.2.
  • Ensemble learning, including bagging-style SVMs and multi-atlas ensembles §3.2.
  • Unsupervised learning: latent class analysis, K-means, PCA, hierarchical clustering, and finite mixture models for longitudinal data such as growth mixture and group-based trajectory models §3.2.
  • Semi-supervised learning: self-training, co-training, multi-view and graph methods, semi-supervised clustering, and normative modelling with Gaussian process regression §3.2 fig. 2.
  • Deep learning: autoencoders, CNNs, graph neural networks, LSTMs and GANs, each with a psychiatric example §3.3.
• Principles for ML systems drawn from an external white paper: trustworthiness, explainability, usability, and transparency with fairness §3.4.
• Digital measurement and intervention §4:
  • Mobile sensing signals, such as location, mobility, call logs, sleep and physiology §4.1 fig. 7.
  • Speech and video features, such as jitter, pause duration, gaze and gait §4.2 tab. 4.
  • Language from EHRs and clinical notes, and from transcripts §4.3.
  • Social media data and the platform access routes to it §4.4 tab. 5.
  • Keystroke, heart-rate and game-behaviour sensors §4.5 tab. 6.
  • Commercial platforms and telehealth §4.6 tab. 7. The patient-journey view of ML use is shown in fig. 6 §4.
• Multimodal fusion §5:
  • Correlation methods: CCA, multiset CCA, PLS and N-PLS §5.1.
  • Factorisation: joint ICA, independent vector analysis, coupled matrix-tensor factorisation, CIFA and JIVE §5.1.
  • Multi-kernel learning, usually with SVMs, and deep fusion §5.1 fig. 8.
  • Neuroimaging fusion studies §5.2 and non-imaging fusion §5.3.
• Molecular phenotyping §6 fig. 9: next-generation sequencing, including bulk and single-nucleus RNA-seq, gene co-expression modules, single-cell tools (Seurat is named), and autoencoders for denoising and transfer.
• Explainable AI and causality §7 fig. 10:
  • Intrinsic versus post-hoc interpretability, rule-based models, constrained models and feedback models §7.1.
  • Circuit-level modelling: Wilson–Cowan neural masses and dynamic causal modelling §7.2.
  • Closed-loop neuromodulation with human-in-the-loop testing §7.3.
• Challenges and opportunities §8.1 §8.2.

The closed loop in fig. 10 is the review's main proposal. The paper gives it as a diagram and in prose, with no algorithm. The sketch below restates that loop; the function names are ours, not the paper's.

# Schematic of the closed loop: neuroimaging -> circuit model ->
# neurostimulation -> behaviour -> revised model  [§7] [fig. 10]
# Schematic only: the paper gives the loop, not these functions.
def closed_loop_round(model, subject, stimulator):
    # 1. Record neuroimaging at baseline  [§7]
    scan = acquire_neuroimaging(subject)
    # 2. Fit a circuit-level model to the scan  [§7.2]
    circuit = fit_circuit_model(model, scan)
    # 3. Apply neurostimulation chosen from the circuit model  [§7.3]
    response = stimulator.apply(circuit.choose_stimulus(subject))
    # 4. Observe the behavioural outcome  [§7]
    behaviour = observe_behaviour(subject, response)
    # 5. Revise the model with the new evidence  [§7]
    return revise_model(model, scan, behaviour)

Hyperparameters

Not applicable. The review reports no training settings of its own. The settings that the cited case studies report are sample sizes, cohort design and cross-validation scheme, and they are listed under Results §3.5.

Results

Author-reported, taken from the cited primary studies. The review does not re-analyse any data. Setup and headline numbers as the review states them:

Other figures the review reports, each as stated by the authors and cited to the work it comes from:

• A 2015 national survey: 61.3% of US psychiatrists use EHRs §4.3.
• An industry figure: telehealth use rose 38-fold relative to before COVID-19 §4.
• A survey of more than 700 psychiatrists: 49% expected ML to help build prognoses, and 54% expected it to help reach diagnoses §4.
• Multi-omics panel with a random forest for PTSD: 85% sensitivity and 77% specificity §6.
• Rodent social-defeat classifier using anxiety and immune measures: 80% sensitivity §6.
• Speech depression screening from a dataset of more than 10,000 speakers, reported by a company §4.2.
• Between 10,000 and 20,000 smartphone apps for mindfulness or CBT §4.6.

Paper vs code

No official code. The paper links no code. Its URLs go to NIMH, CDC, an industry statistics site and a data-centric AI page §1 §8.2. A GitHub repository search for "precision psychiatry" returned 24 repositories, none owned by an author or by the authors' labs. The search for "precision psychiatry machine learning" returned none. This is a review, so no software was expected.

Pitfalls

• Selection, not completeness. The authors say the literature is not exhaustive §1. Their choice of studies drives every claim, so cross-check any claim against its primary source before you rely on it.
• Numbers are secondhand. Most figures come from cited papers, industry reports or websites §1 §4. For example, the economic burden figure and the 38-fold telehealth growth are cited to outside sources, and the review does not check them.
• Table 2 is not comparable. The application table tab. 2 lists method, disease, data type and reference, but gives no metrics. You cannot rank methods from it, and the samples and validation schemes differ across studies.
• Patient-level leakage. The authors warn that studies that do not hold out all data from a given patient can reach wrong conclusions §8.1. Check the validation design of any primary study before you borrow its accuracy.
• Noisy labels. Social media studies often use self-disclosed diagnoses as labels. The review says these labels are noisy and that healthy controls are hard to define §4.4 §8.1.
• Small samples. Neuroimaging datasets in psychiatry are small, and the review says more caution is needed with foundation models in this setting §8.1.
• Interpretability is not trust. Interpretation methods can give erratic results when data are poor or too few, so an explanation does not show the model is right §8.1.
• Causality is open. The closed-loop proposal §7.3 is a research direction. The review names the causal gap but gives no algorithm for choosing stimulation, and it is unclear how to map symptoms onto circuit dimensions §7.3.
• Figures are not used. Several figures reproduce work from other publishers under their own permissions fig. 1 fig. 2 fig. 7. None are carried onto this card; they are cited by number only.
• Metadata fix. The arXiv metadata stored the author "Prathamesh (Param) Kulkarni" split into two entries. This card uses the single name as it appears in the source.
• Version. The metadata records arXiv v2 (updated 2022-07-11). The fetched LaTeX source is the one this card reads.

Lineage

• Builds on: not applicable, as a review. It cites a range of earlier surveys and reviews §1.
• Compared against: the authors contrast their broader scope with narrower reviews §1.
• Superseded by: not stated.