📄 medicine

Identifying antipsychotic treatment episodes using natural language processing and change-point detection

This study presents and validates a scalable framework combining natural language processing and change-point detection to reconstruct longitudinal antipsychotic treatment episodes from unstructured electronic health records, demonstrating clinical validity and revealing significant associations between treatment complexity and healthcare utilization.

Original authors: Tao Wang, Hamilton Morrin, Ninoslav Majkic, David Codling, Risha Govind, Matthew Broadbent, Cecilia Casetta, Giouliana Kadra-Scalzo, Rashmi Patel, Angus Roberts, Richard Dobson, Robert Stewart, Sameer

Published 2026-09-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tao Wang, Hamilton Morrin, Ninoslav Majkic, David Codling, Risha Govind, Matthew Broadbent, Cecilia Casetta, Giouliana Kadra-Scalzo, Rashmi Patel, Angus Roberts, Richard Dobson, Robert Stewart, Sameer Jauhar, James MacCabe

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of mental health care, the story of a patient's treatment is often written in the margins. For decades, doctors have recorded their decisions about medication in free-flowing notes, describing why a drug was started, how a patient responded, or when a switch was made. While these narratives hold the true history of a patient's journey, they are locked away in unstructured text, invisible to the computers that manage hospital data. This creates a blind spot for researchers trying to understand how treatments work over time. They cannot easily see the full picture of who is taking which medicine, for how long, and when the pattern changes. Without this clarity, it is difficult to know if current practices are helping patients or if they are leading to unnecessary hospital stays. The challenge has been to find a way to read these thousands of handwritten notes and turn them into a clear, chronological timeline that a computer can understand and analyze.

A team of researchers at King's College London and the South London and Maudsley NHS Foundation Trust has developed a new method to solve this problem. They created a system that combines two powerful tools: a digital reader that understands human language and a mathematical detector that spots when a pattern changes. The digital reader, known as natural language processing, scans the unstructured clinical notes to find every mention of antipsychotic medication, the drugs used to treat schizophrenia and related conditions. It pulls out the name of the drug, the dose, and the date it was written down. However, simply listing these mentions is not enough, because doctors often write about a medication they stopped years ago or one they are considering for the future. To find the actual periods when a patient was actively taking a specific drug, the team used a change-point detection algorithm. This tool looks at the stream of medication mentions and identifies the exact moments where the pattern shifts, signaling that one treatment has ended and another has begun.

The researchers tested this system on the electronic health records of 12,530 patients diagnosed with schizophrenia-spectrum disorders between 2007 and 2017. They compared the computer's timeline against 113 treatment episodes that had been carefully marked by human experts. The system performed with high accuracy, correctly identifying the start and stop dates of treatment periods in the vast majority of cases. Once validated, the team applied the method to the entire group of patients. They discovered 24,648 distinct treatment episodes, revealing a detailed map of how these patients were treated over a decade. The analysis showed that while many patients stayed on a single medication for a long time, a significant portion experienced complex patterns. In nearly one-third of all treatment episodes, patients were prescribed more than one antipsychotic at the same time. This included both short overlaps, where doctors temporarily gave two drugs while switching from one to another, and longer periods where multiple drugs were used together intentionally.

The study also uncovered how these treatment patterns relate to the patients' need for hospital care. The data showed that patients who were prescribed multiple antipsychotics at the same time, or who went through many different treatment attempts before starting a specific drug called clozapine, were more likely to be admitted to the hospital and stay there for longer periods. Conversely, patients who received long-acting injectable versions of their medication, which are given as a shot rather than a daily pill, tended to use fewer hospital services. The researchers found that the specific type of diagnosis and the patient's background also played a role, with certain groups showing higher rates of hospital use. The ability to automatically reconstruct these long-term treatment histories from messy text offers a new way to monitor the quality of care. It allows health systems to see, in real time, when prescribing habits drift away from guidelines and to identify patients who might benefit from a change in their treatment plan, all without requiring a team of people to read through every single medical note by hand.

Technical Summary: Identifying Antipsychotic Treatment Episodes Using NLP and Change-Point Detection

Problem Statement
Antipsychotic medications are the cornerstone of treatment for schizophrenia-spectrum disorders, yet real-world prescribing often deviates from clinical guidelines, particularly regarding polypharmacy and the timing of clozapine initiation. Systematically characterizing these treatment trajectories is essential for understanding deviations from best practice and their consequences. However, a significant methodological barrier exists: prescribing information is predominantly recorded in unstructured clinical text rather than structured fields. While structured prescription registers exist in some healthcare systems, they capture discrete dispensing events without explicit information on treatment continuity, forcing researchers to rely on arbitrary gap thresholds. Existing approaches to reconstructing treatment episodes from free text are typically labor-intensive, requiring manual chart review, or rely on rule-based thresholds that may not reflect real-world adherence. Furthermore, while Natural Language Processing (NLP) has been applied to extract medication mentions, these methods often capture discrete mentions rather than reconstructing longitudinal trajectories, introducing noise from routine reviews, historical references, or overlapping polypharmacy.

Methodology
The authors propose a scalable framework integrating NLP with Change-Point Detection (CPD) to reconstruct longitudinal antipsychotic treatment episodes from unstructured Electronic Health Records (EHRs). The study utilized data from 12,530 patients with schizophrenia-spectrum disorders (ICD-10 F20–F29) at the South London and Maudsley NHS Foundation Trust (SLaM) between 2007 and 2017.

Data Integration and NLP Extraction: The pipeline combined structured prescription data with unstructured clinical text. A validated NLP pipeline extracted antipsychotic mentions from narrative notes, discharge summaries, and correspondence, mapping synonyms and proprietary names to a standardized pharmacological vocabulary. For each event, the system produced a tuple containing the medication name, formulation (oral or long-acting injectable [LAI]), dose, and the document date.
Time Series Modeling: For each patient, these extracted records were formalized as a discrete categorical time series $S = \{(t_i, m_i)\}$ , where $t_i$ is the timestamp and $m_i$ is the medication variable. To ensure algorithm convergence, the analysis was restricted to patients with at least five records.
Change-Point Detection (CPD): The Pruned Exact Linear Time (PELT) algorithm was applied to detect abrupt shifts in the statistical properties of the time series, identifying points where one treatment replaced another. Unlike rule-based methods, PELT uses dynamic programming with a pruning mechanism to achieve $O(N)$ $O (N)$ expected time complexity without requiring the number of change points to be specified in advance.
- Cost Function: A radial basis function (RBF) kernel was used as the segment-specific cost function to capture non-parametric distributional shifts.
- Hyperparameters: The model was tuned using a penalty parameter ( $\beta$ ) to balance true transition detection against documentation noise, and a min size parameter to enforce a minimum segment length between change points.
Detection of Combined Use: The framework distinguished between two types of combined use:
- Transition Overlap: Temporary overlaps during switching (detected via a 28-day window around changepoints where both drugs were mentioned).
- Concurrent Therapy: Sustained combination therapy (defined as a second antipsychotic mentioned at least three times during an episode or once per month).
Validation: The algorithm was validated against 113 manually annotated episodes from 50 patients. Additionally, a "known-groups" validation compared treatment episode counts between clozapine and non-clozapine users to assess clinical validity at scale. Associations between treatment metrics and healthcare utilization (admissions and inpatient days) were analyzed using Generalized Linear Models (GLMs).

Key Results

Algorithm Performance: The PELT algorithm achieved an F1 score of 0.81 against manual annotations when optimized with a min size of 2 and a penalty of 7. Error analysis identified challenges with medication restarts, sparse records, and transient polypharmacy.
Treatment Episodes: Across 11,731 patients, 24,648 treatment episodes were identified.
Polypharmacy: Combined antipsychotic use was detected in 31.7% of episodes (1,989 transition overlaps and 5,827 concurrent therapy episodes).
Clinical Patterns:
- Clozapine users had a significantly higher mean number of treatment episodes (3.3) compared to non-clozapine users (1.9), consistent with clinical expectations for treatment-resistant schizophrenia.
- Olanzapine was the most common first-line agent, followed by risperidone and aripiprazole. Paliperidone was the most common second-line agent.
- Clozapine had the longest mean episode duration (76.8 months).
Associations with Service Use:
- Higher combination-prescribing rates and a greater number of pre-clozapine episodes were positively associated with increased hospital admissions and longer inpatient stays.
- A higher proportion of LAI episodes relative to total episodes was associated with reduced service use.
- Switching speed (mean episode duration) and the stability of the final episode were not significantly associated with service use.

Significance and Claims
The paper claims to present the first study applying CPD to reconstruct continuous antipsychotic treatment episodes directly from NLP-extracted EHR narratives. The primary significance lies in providing a scalable, automated approach to overcoming the lack of structured prescribing continuity data in mental health EHRs.

The authors assert that this framework:

Overcomes Scalability Limits: It separates genuine clinical transitions from documentation artifacts without the labor-intensive manual abstraction required by previous methods.
Captures Clinical Nuance: It successfully distinguishes between expected cross-titration overlaps and deliberate concurrent therapy, a distinction often lost in rule-based approaches.
Enables Downstream Analysis: The derived metrics (e.g., treatment complexity, transition frequency) can be integrated into statistical models to reveal associations between prescribing patterns and healthcare utilization, such as the link between delayed treatment optimization and increased hospitalization.
Offers Generalizability: The modular architecture, requiring only categorical time-series inputs, is adaptable to other chronic conditions and potential applications in prescribing quality monitoring, automated cohort phenotyping, and clinical decision support systems for identifying treatment resistance.

The authors maintain a modest tone regarding limitations, acknowledging that the pipeline's accuracy depends on upstream NLP extraction, may face challenges in fragmented healthcare networks, and requires recalibration for different documentation styles. They also note that the observational nature of the data precludes causal interpretation of the reported associations.

Technical Summary: Identifying Antipsychotic Treatment Episodes Using NLP and Change-Point Detection

More like this