Navigating promise and perils: applying artificial intelligence to the perinatal mental health care cascade

Abstract

The perinatal mental health care cascade is wrought with systemic issues contributing to under-detection and outcome disparities. Herein, we examine its unique characteristics and explore how artificial intelligence (AI) may improve care while acknowledging associated ethical considerations and implementation challenges. We emphasize the need for policy reforms to screening, data collection, and regulatory processes to build ethical and robust AI-enhanced health system infrastructures.

Keywords: Health care; Information systems and information technology; Psychology.

Fig. 1. Perinatal mental health care cascade and associated barriers.

This figure illustrates the sequential stages of perinatal mental health care delivery and the multilevel barriers that can impede progression through the care continuum. The cascade represents the ideal pathway from initial population screening through achievement of symptom remission, with each stage representing a transition point where patients may be lost to follow-up or experience delays in care. The care cascade consists of four primary stages: (1) Screening & Identification - systematic screening of the perinatal population using validated instruments to identify individuals at risk for or experiencing mental health conditions; (2) Connection to Specialized Care - successful referral and linkage to appropriate mental health services, including initial appointment attendance; (3) Treatment Delivery - provision of evidence-based interventions tailored to perinatal mental health needs; and (4) Remission - achievement of clinically significant symptom reduction and functional improvement. Barriers to care progression are categorized into socioecological domains. Patient-level barriers comprise factors such as mental health stigma, limited social support, and transportation. Provider-level barriers encompass healthcare provider factors, including training and education, resource availability, time constraints, symptom acceptability, and effectiveness. Institutional-level barriers include fragmented clinical networks, services available based on insurance type, staffing levels, and drug shortages. Community barriers encompass system-level obstacles, including the cost of care and support, as well as limited access to specialized providers. Policy-level barriers include systemic inequities, resource allocation, and data gaps.

Fig. 2. Potential AI applications across the PMHC care cascade.

This figure illustrates the various AI applications currently deployed or that could be deployed to enhance the PMHC treatment cascade, as well as the AI systems that underpin these technologies. Clinical applications include four key domains: (1) prediction models and case detection, encompassing risk stratification, early warning systems, and social media monitoring; (2) clinical decision support, providing treatment recommendations, evidence-based guidelines and risk stratification tools; (3) treatment access expansion through teletherapy platforms, AI chatbot support systems, and digital resource navigation; and (4) treatment monitoring via symptom tracking apps, progress assessment tools, and relapse prevention systems. The clinical applications leverage four primary AI technology categories: (a) machine learning, incorporating predictive analytics, classification algorithms, pattern recognition and deep learning; (b) natural language processing enabling text analysis, sentiment analysis, and conversation agents; (c) computer vision systems that include facial expression analysis, behavioral monitoring, and visual interaction tracking; and (d) autonomous AI agents such as virtual assistants, intelligent monitoring systems and adaptive learning agents.

Fig. 3. Potential pitfalls in AI implementation for perinatal mental health.

This figure outlines some of the potential pitfalls of AI implementation across four domains: Bias, including due to the underrepresentation of minoritized groups in training data, limited structured information related to PMHCs outside of depression and anxiety, disparities in screening data resulting in skewed risk assessments, and access inequity in the rollout of systems. Cost implications, including initial set-up costs, employing technical expertise needed for updates, and the cost incurred to monitor models and tailor them to the institution. Privacy and safety concerns outline the risks of breaches, data sharing, consent, and regulatory considerations. Environmental impact encompasses concerns about pollution, the placement of AI infrastructure in marginalized communities, and the natural resources demands required to sustain these systems.