The main problem across all six is the same: there is still no human evidence that any trazodone dose is disease-modifying in dementia. The 2022 cohort critique only shows that real-world exposures were usually low, confounded by indication, and likely late in disease; it does not validate the ISR mechanism or any threshold in patients.[Cambridge repository / Int J Geriatr Psychiatry 2022](https://www.repository.cam.ac.uk/items/a33a4c04-ec4a-4db3-8dd9-a873090dd015) The strongest biological support remains preclinical mouse work on PERK/eIF2alpha-linked translational rescue, which is informative but a weak basis for a precise human dose threshold.[PMID:28430857](https://pubmed.ncbi.nlm.nih.gov/28430857/) [PMID:37703312](https://pubmed.ncbi.nlm.nih.gov/37703312/)
1. **Direct ISR/UPR target engagement requires about 150-200 mg/day**
- Weak links: the `~194 mg/day` figure is an HED extrapolation from mouse efficacy, not a demonstrated human CNS target-engagement threshold. Human dementia brains show ISR activation, but that does not prove trazodone reaches or modulates that node at any given clinical dose.[PMID:12499843](https://pubmed.ncbi.nlm.nih.gov/12499843/) [PMID:28430857](https://pubmed.ncbi.nlm.nih.gov/28430857/)
- Counter-evidence: trazodone’s known human pharmacology is dominated by receptor occupancy at low doses, and the 2018 PK modeling paper itself is inferential, not a biomarker trial.[PMID:29332554](https://pubmed.ncbi.nlm.nih.gov/29332554/)
- Alternative explanation: mouse neuroprotection may reflect mixed sedative/serotonergic/network effects rather than selective ISR rescue in humans.
- Falsifying experiment: a blinded ascending-dose study with CSF or neuron-derived EV evidence showing no monotonic ISR change even at `200 mg/day`, despite adequate drug exposure.
- Revised confidence: `0.42`
2. **A lower threshold, about 50-100 mg nightly, is sufficient only in the sleep-fragmented/SWS-deficient subgroup**
- Weak links: this assumes improving sleep architecture translates into slower neurodegeneration, but the human evidence mostly links sleep disruption to biomarkers or risk, not to reversal of established dementia biology.[PMID:32057125](https://pubmed.ncbi.nlm.nih.gov/32057125/) [PMID:37902739](https://pubmed.ncbi.nlm.nih.gov/37902739/) The glymphatic argument is still partly speculative and measurement is indirect.[PMID:34902819](https://pubmed.ncbi.nlm.nih.gov/34902819/)
- Counter-evidence: the AD sleep RCT with trazodone was short and assessed sleep, not disease modification.[DOI:10.1016/j.jagp.2013.12.174](https://www.sciencedirect.com/science/article/abs/pii/S1064748114000037)
- Alternative explanation: any apparent subgroup benefit could just reflect symptomatic improvement from better sleep, reduced agitation, or caregiver-reported function.
- Falsifying experiment: enrich for low-SWS patients and show that objectively improved N3 does not alter downstream tau/NfL trajectories versus placebo over 6-12 months.
- Revised confidence: `0.33`
3. **The threshold depends more on overnight brain exposure than on total daily dose**
- Weak links: this is pharmacologically plausible but currently almost evidence-free in dementia. The cited support is mostly PK reasoning and animal data, not human CSF-PD evidence.[PMID:29332554](https://pubmed.ncbi.nlm.nih.gov/29332554/) [PMID:28430857](https://pubmed.ncbi.nlm.nih.gov/28430857/)
- Counter-evidence: if trazodone’s putative disease-modifying action depends on cumulative integrated exposure rather than nocturnal peak/trough shape, ER or split dosing may not matter.
- Alternative explanation: regimen differences could affect tolerability and sedation more than neuronal ISR biology.
- Falsifying experiment: matched-AUC crossover with CSF exposure and ISR biomarkers showing no advantage for flatter overnight exposure.
- Revised confidence: `0.25`
4. **Tauopathy biology lowers the effective threshold if p-tau/ISR activation is already high**
- Weak links: it assumes “more activated pathway = more drug sensitivity,” which is not generally true. High p-tau/high ISR may instead mark more advanced, less reversible disease.[PMID:12499843](https://pubmed.ncbi.nlm.nih.gov/12499843/) [PMID:26450683](https://pubmed.ncbi.nlm.nih.gov/26450683/)
- Counter-evidence: in many neurodegenerative settings, severe pathway activation predicts treatment resistance, not enhanced responsiveness.
- Alternative explanation: any signal in high-p-tau patients could reflect regression to the mean or larger biomarker dynamic range rather than lower dose need.
- Falsifying experiment: prestratified trial showing the high-p-tau/high-ISR subgroup fails to respond at `100 mg` despite confirmed exposure.
- Revised confidence: `0.31`
5. **A glial anti-inflammatory threshold exists at about 50-100 mg, separate from the neuronal ISR threshold**
- Weak links: nearly all support is in vitro or cell-line work, including inflammatory paradigms far removed from human dementia brain pharmacology.[PMID:25911310](https://pubmed.ncbi.nlm.nih.gov/25911310/) [PMID:39187397](https://pubmed.ncbi.nlm.nih.gov/39187397/) A glial biomarker shift is not the same as disease modification.
- Counter-evidence: dementia neuroinflammation is heterogeneous and may be epiphenomenal in many patients. Lowering IL-6 or microglial markers may not slow degeneration.
- Alternative explanation: anti-inflammatory readouts may reflect nonspecific stress reduction or off-target cell effects at concentrations not achieved in brain tissue.
- Falsifying experiment: demonstrate no dose-related change in CSF/plasma inflammatory markers at clinically tolerated `50-150 mg/day`, or biomarker change without clinical/structural benefit.
- Revised confidence: `0.28`
6. **The minimum effective dose is higher in dementia patients with OSA/high arousal burden**
- Weak links: this chains together too many unproven steps: OSA physiology -> trazodone-mediated arousal effects -> improved nocturnal pathology -> slower dementia progression. The OSA studies were small, short, and in general OSA populations, not dementia patients.[PMID:18256066](https://pubmed.ncbi.nlm.nih.gov/18256066/) [PMID:25719754](https://pubmed.ncbi.nlm.nih.gov/25719754/)
- Counter-evidence: one OSA study found reduced AHI without a significant arousal-threshold change, which weakens the proposed mechanism.[PMID:25719754](https://pubmed.ncbi.nlm.nih.gov/25719754/)
- Alternative explanation: CPAP adherence, body habitus, airway anatomy, and baseline OSA endotype likely dominate any trazodone effect.
- Falsifying experiment: add trazodone versus placebo on top of optimized OSA care and show no biomarker or sleep-fragmentation advantage at `100 mg`.
- Revised confidence: `0.22`
**Bottom line**
The least weak hypothesis is still `#1`, but only as a **trial-enabling mechanistic prior**, not as a credible human dose estimate. The most defensible near-term model is not “two established thresholds,” but “multiple possible exposure-response regimes that remain unvalidated in humans.” If forced to rank them: `#1 > #2 ≈ #4 > #5 > #3 > #6`.
**Best falsification strategy overall**
Run a biomarker-rich phase Ib/II study that separates:
- symptomatic sleep effects,
- CNS exposure,
- ISR target engagement,
- inflammatory marker shifts,
- and downstream neurodegeneration markers.
Without that separation, any claimed “minimum effective dose” will remain mostly a mechanistic story built on mouse translation and confounded sleep pharmacology.
Sources: [2022 cohort critique](https://www.repository.cam.ac.uk/items/a33a4c04-ec4a-4db3-8dd9-a873090dd015), [PMID:28430857](https://pubmed.ncbi.nlm.nih.gov/28430857/), [PMID:37703312](https://pubmed.ncbi.nlm.nih.gov/37703312/), [PMID:12499843](https://pubmed.ncbi.nlm.nih.gov/12499843/), [PMID:26450683](https://pubmed.ncbi.nlm.nih.gov/26450683/), [PMID:29332554](https://pubmed.ncbi.nlm.nih.gov/29332554/), [PMID:32057125](https://pubmed.ncbi.nlm.nih.gov/32057125/), [PMID:37902739](https://pubmed.ncbi.nlm.nih.gov/37902739/), [PMID:34902819](https://pubmed.ncbi.nlm.nih.gov/34902819/), [PMID:25911310](https://pubmed.ncbi.nlm.nih.gov/25911310/), [PMID:39187397](https://pubmed.ncbi.nlm.nih.gov/39187397/), [PMID:18256066](https://pubmed.ncbi.nlm.nih.gov/18256066/), [PMID:25719754](https://pubmed.ncbi.nlm.nih.gov/25719754/)
If useful, I can convert this into a structured table or JSON with fields like `weak_links`, `counter_evidence`, `falsifier`, and `revised_confidence`.